1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Fast Userspace Mutexes (which I call "Futexes!").
4 * (C) Rusty Russell, IBM 2002
6 * Generalized futexes, futex requeueing, misc fixes by Ingo Molnar
7 * (C) Copyright 2003 Red Hat Inc, All Rights Reserved
9 * Removed page pinning, fix privately mapped COW pages and other cleanups
10 * (C) Copyright 2003, 2004 Jamie Lokier
12 * Robust futex support started by Ingo Molnar
13 * (C) Copyright 2006 Red Hat Inc, All Rights Reserved
14 * Thanks to Thomas Gleixner for suggestions, analysis and fixes.
16 * PI-futex support started by Ingo Molnar and Thomas Gleixner
17 * Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
18 * Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
20 * PRIVATE futexes by Eric Dumazet
21 * Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
23 * Requeue-PI support by Darren Hart <dvhltc@us.ibm.com>
24 * Copyright (C) IBM Corporation, 2009
25 * Thanks to Thomas Gleixner for conceptual design and careful reviews.
27 * Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
28 * enough at me, Linus for the original (flawed) idea, Matthew
29 * Kirkwood for proof-of-concept implementation.
31 * "The futexes are also cursed."
32 * "But they come in a choice of three flavours!"
34 #include <linux/compat.h>
35 #include <linux/jhash.h>
36 #include <linux/pagemap.h>
37 #include <linux/syscalls.h>
38 #include <linux/freezer.h>
39 #include <linux/memblock.h>
40 #include <linux/fault-inject.h>
41 #include <linux/time_namespace.h>
43 #include <asm/futex.h>
45 #include "../locking/rtmutex_common.h"
48 * READ this before attempting to hack on futexes!
50 * Basic futex operation and ordering guarantees
51 * =============================================
53 * The waiter reads the futex value in user space and calls
54 * futex_wait(). This function computes the hash bucket and acquires
55 * the hash bucket lock. After that it reads the futex user space value
56 * again and verifies that the data has not changed. If it has not changed
57 * it enqueues itself into the hash bucket, releases the hash bucket lock
60 * The waker side modifies the user space value of the futex and calls
61 * futex_wake(). This function computes the hash bucket and acquires the
62 * hash bucket lock. Then it looks for waiters on that futex in the hash
63 * bucket and wakes them.
65 * In futex wake up scenarios where no tasks are blocked on a futex, taking
66 * the hb spinlock can be avoided and simply return. In order for this
67 * optimization to work, ordering guarantees must exist so that the waiter
68 * being added to the list is acknowledged when the list is concurrently being
69 * checked by the waker, avoiding scenarios like the following:
73 * sys_futex(WAIT, futex, val);
74 * futex_wait(futex, val);
77 * sys_futex(WAKE, futex);
82 * lock(hash_bucket(futex));
84 * unlock(hash_bucket(futex));
87 * This would cause the waiter on CPU 0 to wait forever because it
88 * missed the transition of the user space value from val to newval
89 * and the waker did not find the waiter in the hash bucket queue.
91 * The correct serialization ensures that a waiter either observes
92 * the changed user space value before blocking or is woken by a
97 * sys_futex(WAIT, futex, val);
98 * futex_wait(futex, val);
101 * smp_mb(); (A) <-- paired with -.
103 * lock(hash_bucket(futex)); |
107 * | sys_futex(WAKE, futex);
108 * | futex_wake(futex);
110 * `--------> smp_mb(); (B)
113 * unlock(hash_bucket(futex));
114 * schedule(); if (waiters)
115 * lock(hash_bucket(futex));
116 * else wake_waiters(futex);
117 * waiters--; (b) unlock(hash_bucket(futex));
119 * Where (A) orders the waiters increment and the futex value read through
120 * atomic operations (see hb_waiters_inc) and where (B) orders the write
121 * to futex and the waiters read (see hb_waiters_pending()).
123 * This yields the following case (where X:=waiters, Y:=futex):
131 * Which guarantees that x==0 && y==0 is impossible; which translates back into
132 * the guarantee that we cannot both miss the futex variable change and the
135 * Note that a new waiter is accounted for in (a) even when it is possible that
136 * the wait call can return error, in which case we backtrack from it in (b).
137 * Refer to the comment in queue_lock().
139 * Similarly, in order to account for waiters being requeued on another
140 * address we always increment the waiters for the destination bucket before
141 * acquiring the lock. It then decrements them again after releasing it -
142 * the code that actually moves the futex(es) between hash buckets (requeue_futex)
143 * will do the additional required waiter count housekeeping. This is done for
144 * double_lock_hb() and double_unlock_hb(), respectively.
147 #ifdef CONFIG_HAVE_FUTEX_CMPXCHG
148 #define futex_cmpxchg_enabled 1
150 static int __read_mostly futex_cmpxchg_enabled;
154 * Futex flags used to encode options to functions and preserve them across
158 # define FLAGS_SHARED 0x01
161 * NOMMU does not have per process address space. Let the compiler optimize
164 # define FLAGS_SHARED 0x00
166 #define FLAGS_CLOCKRT 0x02
167 #define FLAGS_HAS_TIMEOUT 0x04
170 * Priority Inheritance state:
172 struct futex_pi_state {
174 * list of 'owned' pi_state instances - these have to be
175 * cleaned up in do_exit() if the task exits prematurely:
177 struct list_head list;
182 struct rt_mutex pi_mutex;
184 struct task_struct *owner;
188 } __randomize_layout;
191 * struct futex_q - The hashed futex queue entry, one per waiting task
192 * @list: priority-sorted list of tasks waiting on this futex
193 * @task: the task waiting on the futex
194 * @lock_ptr: the hash bucket lock
195 * @key: the key the futex is hashed on
196 * @pi_state: optional priority inheritance state
197 * @rt_waiter: rt_waiter storage for use with requeue_pi
198 * @requeue_pi_key: the requeue_pi target futex key
199 * @bitset: bitset for the optional bitmasked wakeup
201 * We use this hashed waitqueue, instead of a normal wait_queue_entry_t, so
202 * we can wake only the relevant ones (hashed queues may be shared).
204 * A futex_q has a woken state, just like tasks have TASK_RUNNING.
205 * It is considered woken when plist_node_empty(&q->list) || q->lock_ptr == 0.
206 * The order of wakeup is always to make the first condition true, then
209 * PI futexes are typically woken before they are removed from the hash list via
210 * the rt_mutex code. See unqueue_me_pi().
213 struct plist_node list;
215 struct task_struct *task;
216 spinlock_t *lock_ptr;
218 struct futex_pi_state *pi_state;
219 struct rt_mutex_waiter *rt_waiter;
220 union futex_key *requeue_pi_key;
222 } __randomize_layout;
224 static const struct futex_q futex_q_init = {
225 /* list gets initialized in queue_me()*/
226 .key = FUTEX_KEY_INIT,
227 .bitset = FUTEX_BITSET_MATCH_ANY
231 * Hash buckets are shared by all the futex_keys that hash to the same
232 * location. Each key may have multiple futex_q structures, one for each task
233 * waiting on a futex.
235 struct futex_hash_bucket {
238 struct plist_head chain;
239 } ____cacheline_aligned_in_smp;
242 * The base of the bucket array and its size are always used together
243 * (after initialization only in hash_futex()), so ensure that they
244 * reside in the same cacheline.
247 struct futex_hash_bucket *queues;
248 unsigned long hashsize;
249 } __futex_data __read_mostly __aligned(2*sizeof(long));
250 #define futex_queues (__futex_data.queues)
251 #define futex_hashsize (__futex_data.hashsize)
255 * Fault injections for futexes.
257 #ifdef CONFIG_FAIL_FUTEX
260 struct fault_attr attr;
264 .attr = FAULT_ATTR_INITIALIZER,
265 .ignore_private = false,
268 static int __init setup_fail_futex(char *str)
270 return setup_fault_attr(&fail_futex.attr, str);
272 __setup("fail_futex=", setup_fail_futex);
274 static bool should_fail_futex(bool fshared)
276 if (fail_futex.ignore_private && !fshared)
279 return should_fail(&fail_futex.attr, 1);
282 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
284 static int __init fail_futex_debugfs(void)
286 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
289 dir = fault_create_debugfs_attr("fail_futex", NULL,
294 debugfs_create_bool("ignore-private", mode, dir,
295 &fail_futex.ignore_private);
299 late_initcall(fail_futex_debugfs);
301 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
304 static inline bool should_fail_futex(bool fshared)
308 #endif /* CONFIG_FAIL_FUTEX */
311 static void compat_exit_robust_list(struct task_struct *curr);
313 static inline void compat_exit_robust_list(struct task_struct *curr) { }
317 * Reflects a new waiter being added to the waitqueue.
319 static inline void hb_waiters_inc(struct futex_hash_bucket *hb)
322 atomic_inc(&hb->waiters);
324 * Full barrier (A), see the ordering comment above.
326 smp_mb__after_atomic();
331 * Reflects a waiter being removed from the waitqueue by wakeup
334 static inline void hb_waiters_dec(struct futex_hash_bucket *hb)
337 atomic_dec(&hb->waiters);
341 static inline int hb_waiters_pending(struct futex_hash_bucket *hb)
345 * Full barrier (B), see the ordering comment above.
348 return atomic_read(&hb->waiters);
355 * hash_futex - Return the hash bucket in the global hash
356 * @key: Pointer to the futex key for which the hash is calculated
358 * We hash on the keys returned from get_futex_key (see below) and return the
359 * corresponding hash bucket in the global hash.
361 static struct futex_hash_bucket *hash_futex(union futex_key *key)
363 u32 hash = jhash2((u32 *)key, offsetof(typeof(*key), both.offset) / 4,
366 return &futex_queues[hash & (futex_hashsize - 1)];
371 * match_futex - Check whether two futex keys are equal
372 * @key1: Pointer to key1
373 * @key2: Pointer to key2
375 * Return 1 if two futex_keys are equal, 0 otherwise.
377 static inline int match_futex(union futex_key *key1, union futex_key *key2)
380 && key1->both.word == key2->both.word
381 && key1->both.ptr == key2->both.ptr
382 && key1->both.offset == key2->both.offset);
391 * futex_setup_timer - set up the sleeping hrtimer.
392 * @time: ptr to the given timeout value
393 * @timeout: the hrtimer_sleeper structure to be set up
394 * @flags: futex flags
395 * @range_ns: optional range in ns
397 * Return: Initialized hrtimer_sleeper structure or NULL if no timeout
400 static inline struct hrtimer_sleeper *
401 futex_setup_timer(ktime_t *time, struct hrtimer_sleeper *timeout,
402 int flags, u64 range_ns)
407 hrtimer_init_sleeper_on_stack(timeout, (flags & FLAGS_CLOCKRT) ?
408 CLOCK_REALTIME : CLOCK_MONOTONIC,
411 * If range_ns is 0, calling hrtimer_set_expires_range_ns() is
412 * effectively the same as calling hrtimer_set_expires().
414 hrtimer_set_expires_range_ns(&timeout->timer, *time, range_ns);
420 * Generate a machine wide unique identifier for this inode.
422 * This relies on u64 not wrapping in the life-time of the machine; which with
423 * 1ns resolution means almost 585 years.
425 * This further relies on the fact that a well formed program will not unmap
426 * the file while it has a (shared) futex waiting on it. This mapping will have
427 * a file reference which pins the mount and inode.
429 * If for some reason an inode gets evicted and read back in again, it will get
430 * a new sequence number and will _NOT_ match, even though it is the exact same
433 * It is important that match_futex() will never have a false-positive, esp.
434 * for PI futexes that can mess up the state. The above argues that false-negatives
435 * are only possible for malformed programs.
437 static u64 get_inode_sequence_number(struct inode *inode)
439 static atomic64_t i_seq;
442 /* Does the inode already have a sequence number? */
443 old = atomic64_read(&inode->i_sequence);
448 u64 new = atomic64_add_return(1, &i_seq);
449 if (WARN_ON_ONCE(!new))
452 old = atomic64_cmpxchg_relaxed(&inode->i_sequence, 0, new);
460 * get_futex_key() - Get parameters which are the keys for a futex
461 * @uaddr: virtual address of the futex
462 * @fshared: false for a PROCESS_PRIVATE futex, true for PROCESS_SHARED
463 * @key: address where result is stored.
464 * @rw: mapping needs to be read/write (values: FUTEX_READ,
467 * Return: a negative error code or 0
469 * The key words are stored in @key on success.
471 * For shared mappings (when @fshared), the key is:
473 * ( inode->i_sequence, page->index, offset_within_page )
475 * [ also see get_inode_sequence_number() ]
477 * For private mappings (or when !@fshared), the key is:
479 * ( current->mm, address, 0 )
481 * This allows (cross process, where applicable) identification of the futex
482 * without keeping the page pinned for the duration of the FUTEX_WAIT.
484 * lock_page() might sleep, the caller should not hold a spinlock.
486 static int get_futex_key(u32 __user *uaddr, bool fshared, union futex_key *key,
487 enum futex_access rw)
489 unsigned long address = (unsigned long)uaddr;
490 struct mm_struct *mm = current->mm;
491 struct page *page, *tail;
492 struct address_space *mapping;
496 * The futex address must be "naturally" aligned.
498 key->both.offset = address % PAGE_SIZE;
499 if (unlikely((address % sizeof(u32)) != 0))
501 address -= key->both.offset;
503 if (unlikely(!access_ok(uaddr, sizeof(u32))))
506 if (unlikely(should_fail_futex(fshared)))
510 * PROCESS_PRIVATE futexes are fast.
511 * As the mm cannot disappear under us and the 'key' only needs
512 * virtual address, we dont even have to find the underlying vma.
513 * Note : We do have to check 'uaddr' is a valid user address,
514 * but access_ok() should be faster than find_vma()
518 * On no-MMU, shared futexes are treated as private, therefore
519 * we must not include the current process in the key. Since
520 * there is only one address space, the address is a unique key
523 if (IS_ENABLED(CONFIG_MMU))
524 key->private.mm = mm;
526 key->private.mm = NULL;
528 key->private.address = address;
533 /* Ignore any VERIFY_READ mapping (futex common case) */
534 if (unlikely(should_fail_futex(true)))
537 err = get_user_pages_fast(address, 1, FOLL_WRITE, &page);
539 * If write access is not required (eg. FUTEX_WAIT), try
540 * and get read-only access.
542 if (err == -EFAULT && rw == FUTEX_READ) {
543 err = get_user_pages_fast(address, 1, 0, &page);
552 * The treatment of mapping from this point on is critical. The page
553 * lock protects many things but in this context the page lock
554 * stabilizes mapping, prevents inode freeing in the shared
555 * file-backed region case and guards against movement to swap cache.
557 * Strictly speaking the page lock is not needed in all cases being
558 * considered here and page lock forces unnecessarily serialization
559 * From this point on, mapping will be re-verified if necessary and
560 * page lock will be acquired only if it is unavoidable
562 * Mapping checks require the head page for any compound page so the
563 * head page and mapping is looked up now. For anonymous pages, it
564 * does not matter if the page splits in the future as the key is
565 * based on the address. For filesystem-backed pages, the tail is
566 * required as the index of the page determines the key. For
567 * base pages, there is no tail page and tail == page.
570 page = compound_head(page);
571 mapping = READ_ONCE(page->mapping);
574 * If page->mapping is NULL, then it cannot be a PageAnon
575 * page; but it might be the ZERO_PAGE or in the gate area or
576 * in a special mapping (all cases which we are happy to fail);
577 * or it may have been a good file page when get_user_pages_fast
578 * found it, but truncated or holepunched or subjected to
579 * invalidate_complete_page2 before we got the page lock (also
580 * cases which we are happy to fail). And we hold a reference,
581 * so refcount care in invalidate_complete_page's remove_mapping
582 * prevents drop_caches from setting mapping to NULL beneath us.
584 * The case we do have to guard against is when memory pressure made
585 * shmem_writepage move it from filecache to swapcache beneath us:
586 * an unlikely race, but we do need to retry for page->mapping.
588 if (unlikely(!mapping)) {
592 * Page lock is required to identify which special case above
593 * applies. If this is really a shmem page then the page lock
594 * will prevent unexpected transitions.
597 shmem_swizzled = PageSwapCache(page) || page->mapping;
608 * Private mappings are handled in a simple way.
610 * If the futex key is stored on an anonymous page, then the associated
611 * object is the mm which is implicitly pinned by the calling process.
613 * NOTE: When userspace waits on a MAP_SHARED mapping, even if
614 * it's a read-only handle, it's expected that futexes attach to
615 * the object not the particular process.
617 if (PageAnon(page)) {
619 * A RO anonymous page will never change and thus doesn't make
620 * sense for futex operations.
622 if (unlikely(should_fail_futex(true)) || ro) {
627 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
628 key->private.mm = mm;
629 key->private.address = address;
635 * The associated futex object in this case is the inode and
636 * the page->mapping must be traversed. Ordinarily this should
637 * be stabilised under page lock but it's not strictly
638 * necessary in this case as we just want to pin the inode, not
639 * update the radix tree or anything like that.
641 * The RCU read lock is taken as the inode is finally freed
642 * under RCU. If the mapping still matches expectations then the
643 * mapping->host can be safely accessed as being a valid inode.
647 if (READ_ONCE(page->mapping) != mapping) {
654 inode = READ_ONCE(mapping->host);
662 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
663 key->shared.i_seq = get_inode_sequence_number(inode);
664 key->shared.pgoff = page_to_pgoff(tail);
674 * fault_in_user_writeable() - Fault in user address and verify RW access
675 * @uaddr: pointer to faulting user space address
677 * Slow path to fixup the fault we just took in the atomic write
680 * We have no generic implementation of a non-destructive write to the
681 * user address. We know that we faulted in the atomic pagefault
682 * disabled section so we can as well avoid the #PF overhead by
683 * calling get_user_pages() right away.
685 static int fault_in_user_writeable(u32 __user *uaddr)
687 struct mm_struct *mm = current->mm;
691 ret = fixup_user_fault(mm, (unsigned long)uaddr,
692 FAULT_FLAG_WRITE, NULL);
693 mmap_read_unlock(mm);
695 return ret < 0 ? ret : 0;
699 * futex_top_waiter() - Return the highest priority waiter on a futex
700 * @hb: the hash bucket the futex_q's reside in
701 * @key: the futex key (to distinguish it from other futex futex_q's)
703 * Must be called with the hb lock held.
705 static struct futex_q *futex_top_waiter(struct futex_hash_bucket *hb,
706 union futex_key *key)
708 struct futex_q *this;
710 plist_for_each_entry(this, &hb->chain, list) {
711 if (match_futex(&this->key, key))
717 static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
718 u32 uval, u32 newval)
723 ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
729 static int get_futex_value_locked(u32 *dest, u32 __user *from)
734 ret = __get_user(*dest, from);
737 return ret ? -EFAULT : 0;
744 static int refill_pi_state_cache(void)
746 struct futex_pi_state *pi_state;
748 if (likely(current->pi_state_cache))
751 pi_state = kzalloc(sizeof(*pi_state), GFP_KERNEL);
756 INIT_LIST_HEAD(&pi_state->list);
757 /* pi_mutex gets initialized later */
758 pi_state->owner = NULL;
759 refcount_set(&pi_state->refcount, 1);
760 pi_state->key = FUTEX_KEY_INIT;
762 current->pi_state_cache = pi_state;
767 static struct futex_pi_state *alloc_pi_state(void)
769 struct futex_pi_state *pi_state = current->pi_state_cache;
772 current->pi_state_cache = NULL;
777 static void pi_state_update_owner(struct futex_pi_state *pi_state,
778 struct task_struct *new_owner)
780 struct task_struct *old_owner = pi_state->owner;
782 lockdep_assert_held(&pi_state->pi_mutex.wait_lock);
785 raw_spin_lock(&old_owner->pi_lock);
786 WARN_ON(list_empty(&pi_state->list));
787 list_del_init(&pi_state->list);
788 raw_spin_unlock(&old_owner->pi_lock);
792 raw_spin_lock(&new_owner->pi_lock);
793 WARN_ON(!list_empty(&pi_state->list));
794 list_add(&pi_state->list, &new_owner->pi_state_list);
795 pi_state->owner = new_owner;
796 raw_spin_unlock(&new_owner->pi_lock);
800 static void get_pi_state(struct futex_pi_state *pi_state)
802 WARN_ON_ONCE(!refcount_inc_not_zero(&pi_state->refcount));
806 * Drops a reference to the pi_state object and frees or caches it
807 * when the last reference is gone.
809 static void put_pi_state(struct futex_pi_state *pi_state)
814 if (!refcount_dec_and_test(&pi_state->refcount))
818 * If pi_state->owner is NULL, the owner is most probably dying
819 * and has cleaned up the pi_state already
821 if (pi_state->owner) {
824 raw_spin_lock_irqsave(&pi_state->pi_mutex.wait_lock, flags);
825 pi_state_update_owner(pi_state, NULL);
826 rt_mutex_proxy_unlock(&pi_state->pi_mutex);
827 raw_spin_unlock_irqrestore(&pi_state->pi_mutex.wait_lock, flags);
830 if (current->pi_state_cache) {
834 * pi_state->list is already empty.
835 * clear pi_state->owner.
836 * refcount is at 0 - put it back to 1.
838 pi_state->owner = NULL;
839 refcount_set(&pi_state->refcount, 1);
840 current->pi_state_cache = pi_state;
844 #ifdef CONFIG_FUTEX_PI
847 * This task is holding PI mutexes at exit time => bad.
848 * Kernel cleans up PI-state, but userspace is likely hosed.
849 * (Robust-futex cleanup is separate and might save the day for userspace.)
851 static void exit_pi_state_list(struct task_struct *curr)
853 struct list_head *next, *head = &curr->pi_state_list;
854 struct futex_pi_state *pi_state;
855 struct futex_hash_bucket *hb;
856 union futex_key key = FUTEX_KEY_INIT;
858 if (!futex_cmpxchg_enabled)
861 * We are a ZOMBIE and nobody can enqueue itself on
862 * pi_state_list anymore, but we have to be careful
863 * versus waiters unqueueing themselves:
865 raw_spin_lock_irq(&curr->pi_lock);
866 while (!list_empty(head)) {
868 pi_state = list_entry(next, struct futex_pi_state, list);
870 hb = hash_futex(&key);
873 * We can race against put_pi_state() removing itself from the
874 * list (a waiter going away). put_pi_state() will first
875 * decrement the reference count and then modify the list, so
876 * its possible to see the list entry but fail this reference
879 * In that case; drop the locks to let put_pi_state() make
880 * progress and retry the loop.
882 if (!refcount_inc_not_zero(&pi_state->refcount)) {
883 raw_spin_unlock_irq(&curr->pi_lock);
885 raw_spin_lock_irq(&curr->pi_lock);
888 raw_spin_unlock_irq(&curr->pi_lock);
890 spin_lock(&hb->lock);
891 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
892 raw_spin_lock(&curr->pi_lock);
894 * We dropped the pi-lock, so re-check whether this
895 * task still owns the PI-state:
897 if (head->next != next) {
898 /* retain curr->pi_lock for the loop invariant */
899 raw_spin_unlock(&pi_state->pi_mutex.wait_lock);
900 spin_unlock(&hb->lock);
901 put_pi_state(pi_state);
905 WARN_ON(pi_state->owner != curr);
906 WARN_ON(list_empty(&pi_state->list));
907 list_del_init(&pi_state->list);
908 pi_state->owner = NULL;
910 raw_spin_unlock(&curr->pi_lock);
911 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
912 spin_unlock(&hb->lock);
914 rt_mutex_futex_unlock(&pi_state->pi_mutex);
915 put_pi_state(pi_state);
917 raw_spin_lock_irq(&curr->pi_lock);
919 raw_spin_unlock_irq(&curr->pi_lock);
922 static inline void exit_pi_state_list(struct task_struct *curr) { }
926 * We need to check the following states:
928 * Waiter | pi_state | pi->owner | uTID | uODIED | ?
930 * [1] NULL | --- | --- | 0 | 0/1 | Valid
931 * [2] NULL | --- | --- | >0 | 0/1 | Valid
933 * [3] Found | NULL | -- | Any | 0/1 | Invalid
935 * [4] Found | Found | NULL | 0 | 1 | Valid
936 * [5] Found | Found | NULL | >0 | 1 | Invalid
938 * [6] Found | Found | task | 0 | 1 | Valid
940 * [7] Found | Found | NULL | Any | 0 | Invalid
942 * [8] Found | Found | task | ==taskTID | 0/1 | Valid
943 * [9] Found | Found | task | 0 | 0 | Invalid
944 * [10] Found | Found | task | !=taskTID | 0/1 | Invalid
946 * [1] Indicates that the kernel can acquire the futex atomically. We
947 * came here due to a stale FUTEX_WAITERS/FUTEX_OWNER_DIED bit.
949 * [2] Valid, if TID does not belong to a kernel thread. If no matching
950 * thread is found then it indicates that the owner TID has died.
952 * [3] Invalid. The waiter is queued on a non PI futex
954 * [4] Valid state after exit_robust_list(), which sets the user space
955 * value to FUTEX_WAITERS | FUTEX_OWNER_DIED.
957 * [5] The user space value got manipulated between exit_robust_list()
958 * and exit_pi_state_list()
960 * [6] Valid state after exit_pi_state_list() which sets the new owner in
961 * the pi_state but cannot access the user space value.
963 * [7] pi_state->owner can only be NULL when the OWNER_DIED bit is set.
965 * [8] Owner and user space value match
967 * [9] There is no transient state which sets the user space TID to 0
968 * except exit_robust_list(), but this is indicated by the
969 * FUTEX_OWNER_DIED bit. See [4]
971 * [10] There is no transient state which leaves owner and user space
972 * TID out of sync. Except one error case where the kernel is denied
973 * write access to the user address, see fixup_pi_state_owner().
976 * Serialization and lifetime rules:
980 * hb -> futex_q, relation
981 * futex_q -> pi_state, relation
983 * (cannot be raw because hb can contain arbitrary amount
986 * pi_mutex->wait_lock:
990 * (and pi_mutex 'obviously')
994 * p->pi_state_list -> pi_state->list, relation
996 * pi_state->refcount:
1004 * pi_mutex->wait_lock
1010 * Validate that the existing waiter has a pi_state and sanity check
1011 * the pi_state against the user space value. If correct, attach to
1014 static int attach_to_pi_state(u32 __user *uaddr, u32 uval,
1015 struct futex_pi_state *pi_state,
1016 struct futex_pi_state **ps)
1018 pid_t pid = uval & FUTEX_TID_MASK;
1023 * Userspace might have messed up non-PI and PI futexes [3]
1025 if (unlikely(!pi_state))
1029 * We get here with hb->lock held, and having found a
1030 * futex_top_waiter(). This means that futex_lock_pi() of said futex_q
1031 * has dropped the hb->lock in between queue_me() and unqueue_me_pi(),
1032 * which in turn means that futex_lock_pi() still has a reference on
1035 * The waiter holding a reference on @pi_state also protects against
1036 * the unlocked put_pi_state() in futex_unlock_pi(), futex_lock_pi()
1037 * and futex_wait_requeue_pi() as it cannot go to 0 and consequently
1038 * free pi_state before we can take a reference ourselves.
1040 WARN_ON(!refcount_read(&pi_state->refcount));
1043 * Now that we have a pi_state, we can acquire wait_lock
1044 * and do the state validation.
1046 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
1049 * Since {uval, pi_state} is serialized by wait_lock, and our current
1050 * uval was read without holding it, it can have changed. Verify it
1051 * still is what we expect it to be, otherwise retry the entire
1054 if (get_futex_value_locked(&uval2, uaddr))
1061 * Handle the owner died case:
1063 if (uval & FUTEX_OWNER_DIED) {
1065 * exit_pi_state_list sets owner to NULL and wakes the
1066 * topmost waiter. The task which acquires the
1067 * pi_state->rt_mutex will fixup owner.
1069 if (!pi_state->owner) {
1071 * No pi state owner, but the user space TID
1072 * is not 0. Inconsistent state. [5]
1077 * Take a ref on the state and return success. [4]
1083 * If TID is 0, then either the dying owner has not
1084 * yet executed exit_pi_state_list() or some waiter
1085 * acquired the rtmutex in the pi state, but did not
1086 * yet fixup the TID in user space.
1088 * Take a ref on the state and return success. [6]
1094 * If the owner died bit is not set, then the pi_state
1095 * must have an owner. [7]
1097 if (!pi_state->owner)
1102 * Bail out if user space manipulated the futex value. If pi
1103 * state exists then the owner TID must be the same as the
1104 * user space TID. [9/10]
1106 if (pid != task_pid_vnr(pi_state->owner))
1110 get_pi_state(pi_state);
1111 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1128 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1133 * wait_for_owner_exiting - Block until the owner has exited
1134 * @ret: owner's current futex lock status
1135 * @exiting: Pointer to the exiting task
1137 * Caller must hold a refcount on @exiting.
1139 static void wait_for_owner_exiting(int ret, struct task_struct *exiting)
1141 if (ret != -EBUSY) {
1142 WARN_ON_ONCE(exiting);
1146 if (WARN_ON_ONCE(ret == -EBUSY && !exiting))
1149 mutex_lock(&exiting->futex_exit_mutex);
1151 * No point in doing state checking here. If the waiter got here
1152 * while the task was in exec()->exec_futex_release() then it can
1153 * have any FUTEX_STATE_* value when the waiter has acquired the
1154 * mutex. OK, if running, EXITING or DEAD if it reached exit()
1155 * already. Highly unlikely and not a problem. Just one more round
1156 * through the futex maze.
1158 mutex_unlock(&exiting->futex_exit_mutex);
1160 put_task_struct(exiting);
1163 static int handle_exit_race(u32 __user *uaddr, u32 uval,
1164 struct task_struct *tsk)
1169 * If the futex exit state is not yet FUTEX_STATE_DEAD, tell the
1170 * caller that the alleged owner is busy.
1172 if (tsk && tsk->futex_state != FUTEX_STATE_DEAD)
1176 * Reread the user space value to handle the following situation:
1180 * sys_exit() sys_futex()
1181 * do_exit() futex_lock_pi()
1182 * futex_lock_pi_atomic()
1183 * exit_signals(tsk) No waiters:
1184 * tsk->flags |= PF_EXITING; *uaddr == 0x00000PID
1185 * mm_release(tsk) Set waiter bit
1186 * exit_robust_list(tsk) { *uaddr = 0x80000PID;
1187 * Set owner died attach_to_pi_owner() {
1188 * *uaddr = 0xC0000000; tsk = get_task(PID);
1189 * } if (!tsk->flags & PF_EXITING) {
1191 * tsk->futex_state = } else {
1192 * FUTEX_STATE_DEAD; if (tsk->futex_state !=
1195 * return -ESRCH; <--- FAIL
1198 * Returning ESRCH unconditionally is wrong here because the
1199 * user space value has been changed by the exiting task.
1201 * The same logic applies to the case where the exiting task is
1204 if (get_futex_value_locked(&uval2, uaddr))
1207 /* If the user space value has changed, try again. */
1212 * The exiting task did not have a robust list, the robust list was
1213 * corrupted or the user space value in *uaddr is simply bogus.
1214 * Give up and tell user space.
1220 * Lookup the task for the TID provided from user space and attach to
1221 * it after doing proper sanity checks.
1223 static int attach_to_pi_owner(u32 __user *uaddr, u32 uval, union futex_key *key,
1224 struct futex_pi_state **ps,
1225 struct task_struct **exiting)
1227 pid_t pid = uval & FUTEX_TID_MASK;
1228 struct futex_pi_state *pi_state;
1229 struct task_struct *p;
1232 * We are the first waiter - try to look up the real owner and attach
1233 * the new pi_state to it, but bail out when TID = 0 [1]
1235 * The !pid check is paranoid. None of the call sites should end up
1236 * with pid == 0, but better safe than sorry. Let the caller retry
1240 p = find_get_task_by_vpid(pid);
1242 return handle_exit_race(uaddr, uval, NULL);
1244 if (unlikely(p->flags & PF_KTHREAD)) {
1250 * We need to look at the task state to figure out, whether the
1251 * task is exiting. To protect against the change of the task state
1252 * in futex_exit_release(), we do this protected by p->pi_lock:
1254 raw_spin_lock_irq(&p->pi_lock);
1255 if (unlikely(p->futex_state != FUTEX_STATE_OK)) {
1257 * The task is on the way out. When the futex state is
1258 * FUTEX_STATE_DEAD, we know that the task has finished
1261 int ret = handle_exit_race(uaddr, uval, p);
1263 raw_spin_unlock_irq(&p->pi_lock);
1265 * If the owner task is between FUTEX_STATE_EXITING and
1266 * FUTEX_STATE_DEAD then store the task pointer and keep
1267 * the reference on the task struct. The calling code will
1268 * drop all locks, wait for the task to reach
1269 * FUTEX_STATE_DEAD and then drop the refcount. This is
1270 * required to prevent a live lock when the current task
1271 * preempted the exiting task between the two states.
1281 * No existing pi state. First waiter. [2]
1283 * This creates pi_state, we have hb->lock held, this means nothing can
1284 * observe this state, wait_lock is irrelevant.
1286 pi_state = alloc_pi_state();
1289 * Initialize the pi_mutex in locked state and make @p
1292 rt_mutex_init_proxy_locked(&pi_state->pi_mutex, p);
1294 /* Store the key for possible exit cleanups: */
1295 pi_state->key = *key;
1297 WARN_ON(!list_empty(&pi_state->list));
1298 list_add(&pi_state->list, &p->pi_state_list);
1300 * Assignment without holding pi_state->pi_mutex.wait_lock is safe
1301 * because there is no concurrency as the object is not published yet.
1303 pi_state->owner = p;
1304 raw_spin_unlock_irq(&p->pi_lock);
1313 static int lookup_pi_state(u32 __user *uaddr, u32 uval,
1314 struct futex_hash_bucket *hb,
1315 union futex_key *key, struct futex_pi_state **ps,
1316 struct task_struct **exiting)
1318 struct futex_q *top_waiter = futex_top_waiter(hb, key);
1321 * If there is a waiter on that futex, validate it and
1322 * attach to the pi_state when the validation succeeds.
1325 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1328 * We are the first waiter - try to look up the owner based on
1329 * @uval and attach to it.
1331 return attach_to_pi_owner(uaddr, uval, key, ps, exiting);
1334 static int lock_pi_update_atomic(u32 __user *uaddr, u32 uval, u32 newval)
1339 if (unlikely(should_fail_futex(true)))
1342 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1346 /* If user space value changed, let the caller retry */
1347 return curval != uval ? -EAGAIN : 0;
1351 * futex_lock_pi_atomic() - Atomic work required to acquire a pi aware futex
1352 * @uaddr: the pi futex user address
1353 * @hb: the pi futex hash bucket
1354 * @key: the futex key associated with uaddr and hb
1355 * @ps: the pi_state pointer where we store the result of the
1357 * @task: the task to perform the atomic lock work for. This will
1358 * be "current" except in the case of requeue pi.
1359 * @exiting: Pointer to store the task pointer of the owner task
1360 * which is in the middle of exiting
1361 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1364 * - 0 - ready to wait;
1365 * - 1 - acquired the lock;
1368 * The hb->lock and futex_key refs shall be held by the caller.
1370 * @exiting is only set when the return value is -EBUSY. If so, this holds
1371 * a refcount on the exiting task on return and the caller needs to drop it
1372 * after waiting for the exit to complete.
1374 static int futex_lock_pi_atomic(u32 __user *uaddr, struct futex_hash_bucket *hb,
1375 union futex_key *key,
1376 struct futex_pi_state **ps,
1377 struct task_struct *task,
1378 struct task_struct **exiting,
1381 u32 uval, newval, vpid = task_pid_vnr(task);
1382 struct futex_q *top_waiter;
1386 * Read the user space value first so we can validate a few
1387 * things before proceeding further.
1389 if (get_futex_value_locked(&uval, uaddr))
1392 if (unlikely(should_fail_futex(true)))
1398 if ((unlikely((uval & FUTEX_TID_MASK) == vpid)))
1401 if ((unlikely(should_fail_futex(true))))
1405 * Lookup existing state first. If it exists, try to attach to
1408 top_waiter = futex_top_waiter(hb, key);
1410 return attach_to_pi_state(uaddr, uval, top_waiter->pi_state, ps);
1413 * No waiter and user TID is 0. We are here because the
1414 * waiters or the owner died bit is set or called from
1415 * requeue_cmp_pi or for whatever reason something took the
1418 if (!(uval & FUTEX_TID_MASK)) {
1420 * We take over the futex. No other waiters and the user space
1421 * TID is 0. We preserve the owner died bit.
1423 newval = uval & FUTEX_OWNER_DIED;
1426 /* The futex requeue_pi code can enforce the waiters bit */
1428 newval |= FUTEX_WAITERS;
1430 ret = lock_pi_update_atomic(uaddr, uval, newval);
1431 /* If the take over worked, return 1 */
1432 return ret < 0 ? ret : 1;
1436 * First waiter. Set the waiters bit before attaching ourself to
1437 * the owner. If owner tries to unlock, it will be forced into
1438 * the kernel and blocked on hb->lock.
1440 newval = uval | FUTEX_WAITERS;
1441 ret = lock_pi_update_atomic(uaddr, uval, newval);
1445 * If the update of the user space value succeeded, we try to
1446 * attach to the owner. If that fails, no harm done, we only
1447 * set the FUTEX_WAITERS bit in the user space variable.
1449 return attach_to_pi_owner(uaddr, newval, key, ps, exiting);
1453 * __unqueue_futex() - Remove the futex_q from its futex_hash_bucket
1454 * @q: The futex_q to unqueue
1456 * The q->lock_ptr must not be NULL and must be held by the caller.
1458 static void __unqueue_futex(struct futex_q *q)
1460 struct futex_hash_bucket *hb;
1462 if (WARN_ON_SMP(!q->lock_ptr) || WARN_ON(plist_node_empty(&q->list)))
1464 lockdep_assert_held(q->lock_ptr);
1466 hb = container_of(q->lock_ptr, struct futex_hash_bucket, lock);
1467 plist_del(&q->list, &hb->chain);
1472 * The hash bucket lock must be held when this is called.
1473 * Afterwards, the futex_q must not be accessed. Callers
1474 * must ensure to later call wake_up_q() for the actual
1477 static void mark_wake_futex(struct wake_q_head *wake_q, struct futex_q *q)
1479 struct task_struct *p = q->task;
1481 if (WARN(q->pi_state || q->rt_waiter, "refusing to wake PI futex\n"))
1487 * The waiting task can free the futex_q as soon as q->lock_ptr = NULL
1488 * is written, without taking any locks. This is possible in the event
1489 * of a spurious wakeup, for example. A memory barrier is required here
1490 * to prevent the following store to lock_ptr from getting ahead of the
1491 * plist_del in __unqueue_futex().
1493 smp_store_release(&q->lock_ptr, NULL);
1496 * Queue the task for later wakeup for after we've released
1499 wake_q_add_safe(wake_q, p);
1503 * Caller must hold a reference on @pi_state.
1505 static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_pi_state *pi_state)
1508 struct task_struct *new_owner;
1509 bool postunlock = false;
1510 DEFINE_WAKE_Q(wake_q);
1513 new_owner = rt_mutex_next_owner(&pi_state->pi_mutex);
1514 if (WARN_ON_ONCE(!new_owner)) {
1516 * As per the comment in futex_unlock_pi() this should not happen.
1518 * When this happens, give up our locks and try again, giving
1519 * the futex_lock_pi() instance time to complete, either by
1520 * waiting on the rtmutex or removing itself from the futex
1528 * We pass it to the next owner. The WAITERS bit is always kept
1529 * enabled while there is PI state around. We cleanup the owner
1530 * died bit, because we are the owner.
1532 newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
1534 if (unlikely(should_fail_futex(true))) {
1539 ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
1540 if (!ret && (curval != uval)) {
1542 * If a unconditional UNLOCK_PI operation (user space did not
1543 * try the TID->0 transition) raced with a waiter setting the
1544 * FUTEX_WAITERS flag between get_user() and locking the hash
1545 * bucket lock, retry the operation.
1547 if ((FUTEX_TID_MASK & curval) == uval)
1555 * This is a point of no return; once we modified the uval
1556 * there is no going back and subsequent operations must
1559 pi_state_update_owner(pi_state, new_owner);
1560 postunlock = __rt_mutex_futex_unlock(&pi_state->pi_mutex, &wake_q);
1564 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
1567 rt_mutex_postunlock(&wake_q);
1573 * Express the locking dependencies for lockdep:
1576 double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1579 spin_lock(&hb1->lock);
1581 spin_lock_nested(&hb2->lock, SINGLE_DEPTH_NESTING);
1582 } else { /* hb1 > hb2 */
1583 spin_lock(&hb2->lock);
1584 spin_lock_nested(&hb1->lock, SINGLE_DEPTH_NESTING);
1589 double_unlock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
1591 spin_unlock(&hb1->lock);
1593 spin_unlock(&hb2->lock);
1597 * Wake up waiters matching bitset queued on this futex (uaddr).
1600 futex_wake(u32 __user *uaddr, unsigned int flags, int nr_wake, u32 bitset)
1602 struct futex_hash_bucket *hb;
1603 struct futex_q *this, *next;
1604 union futex_key key = FUTEX_KEY_INIT;
1606 DEFINE_WAKE_Q(wake_q);
1611 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_READ);
1612 if (unlikely(ret != 0))
1615 hb = hash_futex(&key);
1617 /* Make sure we really have tasks to wakeup */
1618 if (!hb_waiters_pending(hb))
1621 spin_lock(&hb->lock);
1623 plist_for_each_entry_safe(this, next, &hb->chain, list) {
1624 if (match_futex (&this->key, &key)) {
1625 if (this->pi_state || this->rt_waiter) {
1630 /* Check if one of the bits is set in both bitsets */
1631 if (!(this->bitset & bitset))
1634 mark_wake_futex(&wake_q, this);
1635 if (++ret >= nr_wake)
1640 spin_unlock(&hb->lock);
1645 static int futex_atomic_op_inuser(unsigned int encoded_op, u32 __user *uaddr)
1647 unsigned int op = (encoded_op & 0x70000000) >> 28;
1648 unsigned int cmp = (encoded_op & 0x0f000000) >> 24;
1649 int oparg = sign_extend32((encoded_op & 0x00fff000) >> 12, 11);
1650 int cmparg = sign_extend32(encoded_op & 0x00000fff, 11);
1653 if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28)) {
1654 if (oparg < 0 || oparg > 31) {
1655 char comm[sizeof(current->comm)];
1657 * kill this print and return -EINVAL when userspace
1660 pr_info_ratelimited("futex_wake_op: %s tries to shift op by %d; fix this program\n",
1661 get_task_comm(comm, current), oparg);
1667 pagefault_disable();
1668 ret = arch_futex_atomic_op_inuser(op, oparg, &oldval, uaddr);
1674 case FUTEX_OP_CMP_EQ:
1675 return oldval == cmparg;
1676 case FUTEX_OP_CMP_NE:
1677 return oldval != cmparg;
1678 case FUTEX_OP_CMP_LT:
1679 return oldval < cmparg;
1680 case FUTEX_OP_CMP_GE:
1681 return oldval >= cmparg;
1682 case FUTEX_OP_CMP_LE:
1683 return oldval <= cmparg;
1684 case FUTEX_OP_CMP_GT:
1685 return oldval > cmparg;
1692 * Wake up all waiters hashed on the physical page that is mapped
1693 * to this virtual address:
1696 futex_wake_op(u32 __user *uaddr1, unsigned int flags, u32 __user *uaddr2,
1697 int nr_wake, int nr_wake2, int op)
1699 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1700 struct futex_hash_bucket *hb1, *hb2;
1701 struct futex_q *this, *next;
1703 DEFINE_WAKE_Q(wake_q);
1706 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1707 if (unlikely(ret != 0))
1709 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
1710 if (unlikely(ret != 0))
1713 hb1 = hash_futex(&key1);
1714 hb2 = hash_futex(&key2);
1717 double_lock_hb(hb1, hb2);
1718 op_ret = futex_atomic_op_inuser(op, uaddr2);
1719 if (unlikely(op_ret < 0)) {
1720 double_unlock_hb(hb1, hb2);
1722 if (!IS_ENABLED(CONFIG_MMU) ||
1723 unlikely(op_ret != -EFAULT && op_ret != -EAGAIN)) {
1725 * we don't get EFAULT from MMU faults if we don't have
1726 * an MMU, but we might get them from range checking
1732 if (op_ret == -EFAULT) {
1733 ret = fault_in_user_writeable(uaddr2);
1738 if (!(flags & FLAGS_SHARED)) {
1747 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
1748 if (match_futex (&this->key, &key1)) {
1749 if (this->pi_state || this->rt_waiter) {
1753 mark_wake_futex(&wake_q, this);
1754 if (++ret >= nr_wake)
1761 plist_for_each_entry_safe(this, next, &hb2->chain, list) {
1762 if (match_futex (&this->key, &key2)) {
1763 if (this->pi_state || this->rt_waiter) {
1767 mark_wake_futex(&wake_q, this);
1768 if (++op_ret >= nr_wake2)
1776 double_unlock_hb(hb1, hb2);
1782 * requeue_futex() - Requeue a futex_q from one hb to another
1783 * @q: the futex_q to requeue
1784 * @hb1: the source hash_bucket
1785 * @hb2: the target hash_bucket
1786 * @key2: the new key for the requeued futex_q
1789 void requeue_futex(struct futex_q *q, struct futex_hash_bucket *hb1,
1790 struct futex_hash_bucket *hb2, union futex_key *key2)
1794 * If key1 and key2 hash to the same bucket, no need to
1797 if (likely(&hb1->chain != &hb2->chain)) {
1798 plist_del(&q->list, &hb1->chain);
1799 hb_waiters_dec(hb1);
1800 hb_waiters_inc(hb2);
1801 plist_add(&q->list, &hb2->chain);
1802 q->lock_ptr = &hb2->lock;
1808 * requeue_pi_wake_futex() - Wake a task that acquired the lock during requeue
1810 * @key: the key of the requeue target futex
1811 * @hb: the hash_bucket of the requeue target futex
1813 * During futex_requeue, with requeue_pi=1, it is possible to acquire the
1814 * target futex if it is uncontended or via a lock steal. Set the futex_q key
1815 * to the requeue target futex so the waiter can detect the wakeup on the right
1816 * futex, but remove it from the hb and NULL the rt_waiter so it can detect
1817 * atomic lock acquisition. Set the q->lock_ptr to the requeue target hb->lock
1818 * to protect access to the pi_state to fixup the owner later. Must be called
1819 * with both q->lock_ptr and hb->lock held.
1822 void requeue_pi_wake_futex(struct futex_q *q, union futex_key *key,
1823 struct futex_hash_bucket *hb)
1829 WARN_ON(!q->rt_waiter);
1830 q->rt_waiter = NULL;
1832 q->lock_ptr = &hb->lock;
1834 wake_up_state(q->task, TASK_NORMAL);
1838 * futex_proxy_trylock_atomic() - Attempt an atomic lock for the top waiter
1839 * @pifutex: the user address of the to futex
1840 * @hb1: the from futex hash bucket, must be locked by the caller
1841 * @hb2: the to futex hash bucket, must be locked by the caller
1842 * @key1: the from futex key
1843 * @key2: the to futex key
1844 * @ps: address to store the pi_state pointer
1845 * @exiting: Pointer to store the task pointer of the owner task
1846 * which is in the middle of exiting
1847 * @set_waiters: force setting the FUTEX_WAITERS bit (1) or not (0)
1849 * Try and get the lock on behalf of the top waiter if we can do it atomically.
1850 * Wake the top waiter if we succeed. If the caller specified set_waiters,
1851 * then direct futex_lock_pi_atomic() to force setting the FUTEX_WAITERS bit.
1852 * hb1 and hb2 must be held by the caller.
1854 * @exiting is only set when the return value is -EBUSY. If so, this holds
1855 * a refcount on the exiting task on return and the caller needs to drop it
1856 * after waiting for the exit to complete.
1859 * - 0 - failed to acquire the lock atomically;
1860 * - >0 - acquired the lock, return value is vpid of the top_waiter
1864 futex_proxy_trylock_atomic(u32 __user *pifutex, struct futex_hash_bucket *hb1,
1865 struct futex_hash_bucket *hb2, union futex_key *key1,
1866 union futex_key *key2, struct futex_pi_state **ps,
1867 struct task_struct **exiting, int set_waiters)
1869 struct futex_q *top_waiter = NULL;
1873 if (get_futex_value_locked(&curval, pifutex))
1876 if (unlikely(should_fail_futex(true)))
1880 * Find the top_waiter and determine if there are additional waiters.
1881 * If the caller intends to requeue more than 1 waiter to pifutex,
1882 * force futex_lock_pi_atomic() to set the FUTEX_WAITERS bit now,
1883 * as we have means to handle the possible fault. If not, don't set
1884 * the bit unecessarily as it will force the subsequent unlock to enter
1887 top_waiter = futex_top_waiter(hb1, key1);
1889 /* There are no waiters, nothing for us to do. */
1893 /* Ensure we requeue to the expected futex. */
1894 if (!match_futex(top_waiter->requeue_pi_key, key2))
1898 * Try to take the lock for top_waiter. Set the FUTEX_WAITERS bit in
1899 * the contended case or if set_waiters is 1. The pi_state is returned
1900 * in ps in contended cases.
1902 vpid = task_pid_vnr(top_waiter->task);
1903 ret = futex_lock_pi_atomic(pifutex, hb2, key2, ps, top_waiter->task,
1904 exiting, set_waiters);
1906 requeue_pi_wake_futex(top_waiter, key2, hb2);
1913 * futex_requeue() - Requeue waiters from uaddr1 to uaddr2
1914 * @uaddr1: source futex user address
1915 * @flags: futex flags (FLAGS_SHARED, etc.)
1916 * @uaddr2: target futex user address
1917 * @nr_wake: number of waiters to wake (must be 1 for requeue_pi)
1918 * @nr_requeue: number of waiters to requeue (0-INT_MAX)
1919 * @cmpval: @uaddr1 expected value (or %NULL)
1920 * @requeue_pi: if we are attempting to requeue from a non-pi futex to a
1921 * pi futex (pi to pi requeue is not supported)
1923 * Requeue waiters on uaddr1 to uaddr2. In the requeue_pi case, try to acquire
1924 * uaddr2 atomically on behalf of the top waiter.
1927 * - >=0 - on success, the number of tasks requeued or woken;
1930 static int futex_requeue(u32 __user *uaddr1, unsigned int flags,
1931 u32 __user *uaddr2, int nr_wake, int nr_requeue,
1932 u32 *cmpval, int requeue_pi)
1934 union futex_key key1 = FUTEX_KEY_INIT, key2 = FUTEX_KEY_INIT;
1935 int task_count = 0, ret;
1936 struct futex_pi_state *pi_state = NULL;
1937 struct futex_hash_bucket *hb1, *hb2;
1938 struct futex_q *this, *next;
1939 DEFINE_WAKE_Q(wake_q);
1941 if (nr_wake < 0 || nr_requeue < 0)
1945 * When PI not supported: return -ENOSYS if requeue_pi is true,
1946 * consequently the compiler knows requeue_pi is always false past
1947 * this point which will optimize away all the conditional code
1950 if (!IS_ENABLED(CONFIG_FUTEX_PI) && requeue_pi)
1955 * Requeue PI only works on two distinct uaddrs. This
1956 * check is only valid for private futexes. See below.
1958 if (uaddr1 == uaddr2)
1962 * requeue_pi requires a pi_state, try to allocate it now
1963 * without any locks in case it fails.
1965 if (refill_pi_state_cache())
1968 * requeue_pi must wake as many tasks as it can, up to nr_wake
1969 * + nr_requeue, since it acquires the rt_mutex prior to
1970 * returning to userspace, so as to not leave the rt_mutex with
1971 * waiters and no owner. However, second and third wake-ups
1972 * cannot be predicted as they involve race conditions with the
1973 * first wake and a fault while looking up the pi_state. Both
1974 * pthread_cond_signal() and pthread_cond_broadcast() should
1982 ret = get_futex_key(uaddr1, flags & FLAGS_SHARED, &key1, FUTEX_READ);
1983 if (unlikely(ret != 0))
1985 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2,
1986 requeue_pi ? FUTEX_WRITE : FUTEX_READ);
1987 if (unlikely(ret != 0))
1991 * The check above which compares uaddrs is not sufficient for
1992 * shared futexes. We need to compare the keys:
1994 if (requeue_pi && match_futex(&key1, &key2))
1997 hb1 = hash_futex(&key1);
1998 hb2 = hash_futex(&key2);
2001 hb_waiters_inc(hb2);
2002 double_lock_hb(hb1, hb2);
2004 if (likely(cmpval != NULL)) {
2007 ret = get_futex_value_locked(&curval, uaddr1);
2009 if (unlikely(ret)) {
2010 double_unlock_hb(hb1, hb2);
2011 hb_waiters_dec(hb2);
2013 ret = get_user(curval, uaddr1);
2017 if (!(flags & FLAGS_SHARED))
2022 if (curval != *cmpval) {
2028 if (requeue_pi && (task_count - nr_wake < nr_requeue)) {
2029 struct task_struct *exiting = NULL;
2032 * Attempt to acquire uaddr2 and wake the top waiter. If we
2033 * intend to requeue waiters, force setting the FUTEX_WAITERS
2034 * bit. We force this here where we are able to easily handle
2035 * faults rather in the requeue loop below.
2037 ret = futex_proxy_trylock_atomic(uaddr2, hb1, hb2, &key1,
2039 &exiting, nr_requeue);
2042 * At this point the top_waiter has either taken uaddr2 or is
2043 * waiting on it. If the former, then the pi_state will not
2044 * exist yet, look it up one more time to ensure we have a
2045 * reference to it. If the lock was taken, ret contains the
2046 * vpid of the top waiter task.
2047 * If the lock was not taken, we have pi_state and an initial
2048 * refcount on it. In case of an error we have nothing.
2054 * If we acquired the lock, then the user space value
2055 * of uaddr2 should be vpid. It cannot be changed by
2056 * the top waiter as it is blocked on hb2 lock if it
2057 * tries to do so. If something fiddled with it behind
2058 * our back the pi state lookup might unearth it. So
2059 * we rather use the known value than rereading and
2060 * handing potential crap to lookup_pi_state.
2062 * If that call succeeds then we have pi_state and an
2063 * initial refcount on it.
2065 ret = lookup_pi_state(uaddr2, ret, hb2, &key2,
2066 &pi_state, &exiting);
2071 /* We hold a reference on the pi state. */
2074 /* If the above failed, then pi_state is NULL */
2076 double_unlock_hb(hb1, hb2);
2077 hb_waiters_dec(hb2);
2078 ret = fault_in_user_writeable(uaddr2);
2085 * Two reasons for this:
2086 * - EBUSY: Owner is exiting and we just wait for the
2088 * - EAGAIN: The user space value changed.
2090 double_unlock_hb(hb1, hb2);
2091 hb_waiters_dec(hb2);
2093 * Handle the case where the owner is in the middle of
2094 * exiting. Wait for the exit to complete otherwise
2095 * this task might loop forever, aka. live lock.
2097 wait_for_owner_exiting(ret, exiting);
2105 plist_for_each_entry_safe(this, next, &hb1->chain, list) {
2106 if (task_count - nr_wake >= nr_requeue)
2109 if (!match_futex(&this->key, &key1))
2113 * FUTEX_WAIT_REQEUE_PI and FUTEX_CMP_REQUEUE_PI should always
2114 * be paired with each other and no other futex ops.
2116 * We should never be requeueing a futex_q with a pi_state,
2117 * which is awaiting a futex_unlock_pi().
2119 if ((requeue_pi && !this->rt_waiter) ||
2120 (!requeue_pi && this->rt_waiter) ||
2127 * Wake nr_wake waiters. For requeue_pi, if we acquired the
2128 * lock, we already woke the top_waiter. If not, it will be
2129 * woken by futex_unlock_pi().
2131 if (++task_count <= nr_wake && !requeue_pi) {
2132 mark_wake_futex(&wake_q, this);
2136 /* Ensure we requeue to the expected futex for requeue_pi. */
2137 if (requeue_pi && !match_futex(this->requeue_pi_key, &key2)) {
2143 * Requeue nr_requeue waiters and possibly one more in the case
2144 * of requeue_pi if we couldn't acquire the lock atomically.
2148 * Prepare the waiter to take the rt_mutex. Take a
2149 * refcount on the pi_state and store the pointer in
2150 * the futex_q object of the waiter.
2152 get_pi_state(pi_state);
2153 this->pi_state = pi_state;
2154 ret = rt_mutex_start_proxy_lock(&pi_state->pi_mutex,
2159 * We got the lock. We do neither drop the
2160 * refcount on pi_state nor clear
2161 * this->pi_state because the waiter needs the
2162 * pi_state for cleaning up the user space
2163 * value. It will drop the refcount after
2166 requeue_pi_wake_futex(this, &key2, hb2);
2170 * rt_mutex_start_proxy_lock() detected a
2171 * potential deadlock when we tried to queue
2172 * that waiter. Drop the pi_state reference
2173 * which we took above and remove the pointer
2174 * to the state from the waiters futex_q
2177 this->pi_state = NULL;
2178 put_pi_state(pi_state);
2180 * We stop queueing more waiters and let user
2181 * space deal with the mess.
2186 requeue_futex(this, hb1, hb2, &key2);
2190 * We took an extra initial reference to the pi_state either
2191 * in futex_proxy_trylock_atomic() or in lookup_pi_state(). We
2192 * need to drop it here again.
2194 put_pi_state(pi_state);
2197 double_unlock_hb(hb1, hb2);
2199 hb_waiters_dec(hb2);
2200 return ret ? ret : task_count;
2203 /* The key must be already stored in q->key. */
2204 static inline struct futex_hash_bucket *queue_lock(struct futex_q *q)
2205 __acquires(&hb->lock)
2207 struct futex_hash_bucket *hb;
2209 hb = hash_futex(&q->key);
2212 * Increment the counter before taking the lock so that
2213 * a potential waker won't miss a to-be-slept task that is
2214 * waiting for the spinlock. This is safe as all queue_lock()
2215 * users end up calling queue_me(). Similarly, for housekeeping,
2216 * decrement the counter at queue_unlock() when some error has
2217 * occurred and we don't end up adding the task to the list.
2219 hb_waiters_inc(hb); /* implies smp_mb(); (A) */
2221 q->lock_ptr = &hb->lock;
2223 spin_lock(&hb->lock);
2228 queue_unlock(struct futex_hash_bucket *hb)
2229 __releases(&hb->lock)
2231 spin_unlock(&hb->lock);
2235 static inline void __queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2240 * The priority used to register this element is
2241 * - either the real thread-priority for the real-time threads
2242 * (i.e. threads with a priority lower than MAX_RT_PRIO)
2243 * - or MAX_RT_PRIO for non-RT threads.
2244 * Thus, all RT-threads are woken first in priority order, and
2245 * the others are woken last, in FIFO order.
2247 prio = min(current->normal_prio, MAX_RT_PRIO);
2249 plist_node_init(&q->list, prio);
2250 plist_add(&q->list, &hb->chain);
2255 * queue_me() - Enqueue the futex_q on the futex_hash_bucket
2256 * @q: The futex_q to enqueue
2257 * @hb: The destination hash bucket
2259 * The hb->lock must be held by the caller, and is released here. A call to
2260 * queue_me() is typically paired with exactly one call to unqueue_me(). The
2261 * exceptions involve the PI related operations, which may use unqueue_me_pi()
2262 * or nothing if the unqueue is done as part of the wake process and the unqueue
2263 * state is implicit in the state of woken task (see futex_wait_requeue_pi() for
2266 static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
2267 __releases(&hb->lock)
2270 spin_unlock(&hb->lock);
2274 * unqueue_me() - Remove the futex_q from its futex_hash_bucket
2275 * @q: The futex_q to unqueue
2277 * The q->lock_ptr must not be held by the caller. A call to unqueue_me() must
2278 * be paired with exactly one earlier call to queue_me().
2281 * - 1 - if the futex_q was still queued (and we removed unqueued it);
2282 * - 0 - if the futex_q was already removed by the waking thread
2284 static int unqueue_me(struct futex_q *q)
2286 spinlock_t *lock_ptr;
2289 /* In the common case we don't take the spinlock, which is nice. */
2292 * q->lock_ptr can change between this read and the following spin_lock.
2293 * Use READ_ONCE to forbid the compiler from reloading q->lock_ptr and
2294 * optimizing lock_ptr out of the logic below.
2296 lock_ptr = READ_ONCE(q->lock_ptr);
2297 if (lock_ptr != NULL) {
2298 spin_lock(lock_ptr);
2300 * q->lock_ptr can change between reading it and
2301 * spin_lock(), causing us to take the wrong lock. This
2302 * corrects the race condition.
2304 * Reasoning goes like this: if we have the wrong lock,
2305 * q->lock_ptr must have changed (maybe several times)
2306 * between reading it and the spin_lock(). It can
2307 * change again after the spin_lock() but only if it was
2308 * already changed before the spin_lock(). It cannot,
2309 * however, change back to the original value. Therefore
2310 * we can detect whether we acquired the correct lock.
2312 if (unlikely(lock_ptr != q->lock_ptr)) {
2313 spin_unlock(lock_ptr);
2318 BUG_ON(q->pi_state);
2320 spin_unlock(lock_ptr);
2328 * PI futexes can not be requeued and must remove themself from the
2329 * hash bucket. The hash bucket lock (i.e. lock_ptr) is held on entry
2332 static void unqueue_me_pi(struct futex_q *q)
2333 __releases(q->lock_ptr)
2337 BUG_ON(!q->pi_state);
2338 put_pi_state(q->pi_state);
2341 spin_unlock(q->lock_ptr);
2344 static int __fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2345 struct task_struct *argowner)
2347 struct futex_pi_state *pi_state = q->pi_state;
2348 struct task_struct *oldowner, *newowner;
2349 u32 uval, curval, newval, newtid;
2352 oldowner = pi_state->owner;
2355 * We are here because either:
2357 * - we stole the lock and pi_state->owner needs updating to reflect
2358 * that (@argowner == current),
2362 * - someone stole our lock and we need to fix things to point to the
2363 * new owner (@argowner == NULL).
2365 * Either way, we have to replace the TID in the user space variable.
2366 * This must be atomic as we have to preserve the owner died bit here.
2368 * Note: We write the user space value _before_ changing the pi_state
2369 * because we can fault here. Imagine swapped out pages or a fork
2370 * that marked all the anonymous memory readonly for cow.
2372 * Modifying pi_state _before_ the user space value would leave the
2373 * pi_state in an inconsistent state when we fault here, because we
2374 * need to drop the locks to handle the fault. This might be observed
2375 * in the PID check in lookup_pi_state.
2379 if (oldowner != current) {
2381 * We raced against a concurrent self; things are
2382 * already fixed up. Nothing to do.
2387 if (__rt_mutex_futex_trylock(&pi_state->pi_mutex)) {
2388 /* We got the lock. pi_state is correct. Tell caller. */
2393 * The trylock just failed, so either there is an owner or
2394 * there is a higher priority waiter than this one.
2396 newowner = rt_mutex_owner(&pi_state->pi_mutex);
2398 * If the higher priority waiter has not yet taken over the
2399 * rtmutex then newowner is NULL. We can't return here with
2400 * that state because it's inconsistent vs. the user space
2401 * state. So drop the locks and try again. It's a valid
2402 * situation and not any different from the other retry
2405 if (unlikely(!newowner)) {
2410 WARN_ON_ONCE(argowner != current);
2411 if (oldowner == current) {
2413 * We raced against a concurrent self; things are
2414 * already fixed up. Nothing to do.
2418 newowner = argowner;
2421 newtid = task_pid_vnr(newowner) | FUTEX_WAITERS;
2423 if (!pi_state->owner)
2424 newtid |= FUTEX_OWNER_DIED;
2426 err = get_futex_value_locked(&uval, uaddr);
2431 newval = (uval & FUTEX_OWNER_DIED) | newtid;
2433 err = cmpxchg_futex_value_locked(&curval, uaddr, uval, newval);
2443 * We fixed up user space. Now we need to fix the pi_state
2446 pi_state_update_owner(pi_state, newowner);
2448 return argowner == current;
2451 * In order to reschedule or handle a page fault, we need to drop the
2452 * locks here. In the case of a fault, this gives the other task
2453 * (either the highest priority waiter itself or the task which stole
2454 * the rtmutex) the chance to try the fixup of the pi_state. So once we
2455 * are back from handling the fault we need to check the pi_state after
2456 * reacquiring the locks and before trying to do another fixup. When
2457 * the fixup has been done already we simply return.
2459 * Note: we hold both hb->lock and pi_mutex->wait_lock. We can safely
2460 * drop hb->lock since the caller owns the hb -> futex_q relation.
2461 * Dropping the pi_mutex->wait_lock requires the state revalidate.
2464 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2465 spin_unlock(q->lock_ptr);
2469 err = fault_in_user_writeable(uaddr);
2482 spin_lock(q->lock_ptr);
2483 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2486 * Check if someone else fixed it for us:
2488 if (pi_state->owner != oldowner)
2489 return argowner == current;
2491 /* Retry if err was -EAGAIN or the fault in succeeded */
2496 * fault_in_user_writeable() failed so user state is immutable. At
2497 * best we can make the kernel state consistent but user state will
2498 * be most likely hosed and any subsequent unlock operation will be
2499 * rejected due to PI futex rule [10].
2501 * Ensure that the rtmutex owner is also the pi_state owner despite
2502 * the user space value claiming something different. There is no
2503 * point in unlocking the rtmutex if current is the owner as it
2504 * would need to wait until the next waiter has taken the rtmutex
2505 * to guarantee consistent state. Keep it simple. Userspace asked
2506 * for this wreckaged state.
2508 * The rtmutex has an owner - either current or some other
2509 * task. See the EAGAIN loop above.
2511 pi_state_update_owner(pi_state, rt_mutex_owner(&pi_state->pi_mutex));
2516 static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
2517 struct task_struct *argowner)
2519 struct futex_pi_state *pi_state = q->pi_state;
2522 lockdep_assert_held(q->lock_ptr);
2524 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
2525 ret = __fixup_pi_state_owner(uaddr, q, argowner);
2526 raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
2530 static long futex_wait_restart(struct restart_block *restart);
2533 * fixup_owner() - Post lock pi_state and corner case management
2534 * @uaddr: user address of the futex
2535 * @q: futex_q (contains pi_state and access to the rt_mutex)
2536 * @locked: if the attempt to take the rt_mutex succeeded (1) or not (0)
2538 * After attempting to lock an rt_mutex, this function is called to cleanup
2539 * the pi_state owner as well as handle race conditions that may allow us to
2540 * acquire the lock. Must be called with the hb lock held.
2543 * - 1 - success, lock taken;
2544 * - 0 - success, lock not taken;
2545 * - <0 - on error (-EFAULT)
2547 static int fixup_owner(u32 __user *uaddr, struct futex_q *q, int locked)
2551 * Got the lock. We might not be the anticipated owner if we
2552 * did a lock-steal - fix up the PI-state in that case:
2554 * Speculative pi_state->owner read (we don't hold wait_lock);
2555 * since we own the lock pi_state->owner == current is the
2556 * stable state, anything else needs more attention.
2558 if (q->pi_state->owner != current)
2559 return fixup_pi_state_owner(uaddr, q, current);
2564 * If we didn't get the lock; check if anybody stole it from us. In
2565 * that case, we need to fix up the uval to point to them instead of
2566 * us, otherwise bad things happen. [10]
2568 * Another speculative read; pi_state->owner == current is unstable
2569 * but needs our attention.
2571 if (q->pi_state->owner == current)
2572 return fixup_pi_state_owner(uaddr, q, NULL);
2575 * Paranoia check. If we did not take the lock, then we should not be
2576 * the owner of the rt_mutex. Warn and establish consistent state.
2578 if (WARN_ON_ONCE(rt_mutex_owner(&q->pi_state->pi_mutex) == current))
2579 return fixup_pi_state_owner(uaddr, q, current);
2585 * futex_wait_queue_me() - queue_me() and wait for wakeup, timeout, or signal
2586 * @hb: the futex hash bucket, must be locked by the caller
2587 * @q: the futex_q to queue up on
2588 * @timeout: the prepared hrtimer_sleeper, or null for no timeout
2590 static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
2591 struct hrtimer_sleeper *timeout)
2594 * The task state is guaranteed to be set before another task can
2595 * wake it. set_current_state() is implemented using smp_store_mb() and
2596 * queue_me() calls spin_unlock() upon completion, both serializing
2597 * access to the hash list and forcing another memory barrier.
2599 set_current_state(TASK_INTERRUPTIBLE);
2604 hrtimer_sleeper_start_expires(timeout, HRTIMER_MODE_ABS);
2607 * If we have been removed from the hash list, then another task
2608 * has tried to wake us, and we can skip the call to schedule().
2610 if (likely(!plist_node_empty(&q->list))) {
2612 * If the timer has already expired, current will already be
2613 * flagged for rescheduling. Only call schedule if there
2614 * is no timeout, or if it has yet to expire.
2616 if (!timeout || timeout->task)
2617 freezable_schedule();
2619 __set_current_state(TASK_RUNNING);
2623 * futex_wait_setup() - Prepare to wait on a futex
2624 * @uaddr: the futex userspace address
2625 * @val: the expected value
2626 * @flags: futex flags (FLAGS_SHARED, etc.)
2627 * @q: the associated futex_q
2628 * @hb: storage for hash_bucket pointer to be returned to caller
2630 * Setup the futex_q and locate the hash_bucket. Get the futex value and
2631 * compare it with the expected value. Handle atomic faults internally.
2632 * Return with the hb lock held and a q.key reference on success, and unlocked
2633 * with no q.key reference on failure.
2636 * - 0 - uaddr contains val and hb has been locked;
2637 * - <1 - -EFAULT or -EWOULDBLOCK (uaddr does not contain val) and hb is unlocked
2639 static int futex_wait_setup(u32 __user *uaddr, u32 val, unsigned int flags,
2640 struct futex_q *q, struct futex_hash_bucket **hb)
2646 * Access the page AFTER the hash-bucket is locked.
2647 * Order is important:
2649 * Userspace waiter: val = var; if (cond(val)) futex_wait(&var, val);
2650 * Userspace waker: if (cond(var)) { var = new; futex_wake(&var); }
2652 * The basic logical guarantee of a futex is that it blocks ONLY
2653 * if cond(var) is known to be true at the time of blocking, for
2654 * any cond. If we locked the hash-bucket after testing *uaddr, that
2655 * would open a race condition where we could block indefinitely with
2656 * cond(var) false, which would violate the guarantee.
2658 * On the other hand, we insert q and release the hash-bucket only
2659 * after testing *uaddr. This guarantees that futex_wait() will NOT
2660 * absorb a wakeup if *uaddr does not match the desired values
2661 * while the syscall executes.
2664 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q->key, FUTEX_READ);
2665 if (unlikely(ret != 0))
2669 *hb = queue_lock(q);
2671 ret = get_futex_value_locked(&uval, uaddr);
2676 ret = get_user(uval, uaddr);
2680 if (!(flags & FLAGS_SHARED))
2694 static int futex_wait(u32 __user *uaddr, unsigned int flags, u32 val,
2695 ktime_t *abs_time, u32 bitset)
2697 struct hrtimer_sleeper timeout, *to;
2698 struct restart_block *restart;
2699 struct futex_hash_bucket *hb;
2700 struct futex_q q = futex_q_init;
2707 to = futex_setup_timer(abs_time, &timeout, flags,
2708 current->timer_slack_ns);
2711 * Prepare to wait on uaddr. On success, holds hb lock and increments
2714 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
2718 /* queue_me and wait for wakeup, timeout, or a signal. */
2719 futex_wait_queue_me(hb, &q, to);
2721 /* If we were woken (and unqueued), we succeeded, whatever. */
2723 /* unqueue_me() drops q.key ref */
2724 if (!unqueue_me(&q))
2727 if (to && !to->task)
2731 * We expect signal_pending(current), but we might be the
2732 * victim of a spurious wakeup as well.
2734 if (!signal_pending(current))
2741 restart = ¤t->restart_block;
2742 restart->futex.uaddr = uaddr;
2743 restart->futex.val = val;
2744 restart->futex.time = *abs_time;
2745 restart->futex.bitset = bitset;
2746 restart->futex.flags = flags | FLAGS_HAS_TIMEOUT;
2748 ret = set_restart_fn(restart, futex_wait_restart);
2752 hrtimer_cancel(&to->timer);
2753 destroy_hrtimer_on_stack(&to->timer);
2759 static long futex_wait_restart(struct restart_block *restart)
2761 u32 __user *uaddr = restart->futex.uaddr;
2762 ktime_t t, *tp = NULL;
2764 if (restart->futex.flags & FLAGS_HAS_TIMEOUT) {
2765 t = restart->futex.time;
2768 restart->fn = do_no_restart_syscall;
2770 return (long)futex_wait(uaddr, restart->futex.flags,
2771 restart->futex.val, tp, restart->futex.bitset);
2776 * Userspace tried a 0 -> TID atomic transition of the futex value
2777 * and failed. The kernel side here does the whole locking operation:
2778 * if there are waiters then it will block as a consequence of relying
2779 * on rt-mutexes, it does PI, etc. (Due to races the kernel might see
2780 * a 0 value of the futex too.).
2782 * Also serves as futex trylock_pi()'ing, and due semantics.
2784 static int futex_lock_pi(u32 __user *uaddr, unsigned int flags,
2785 ktime_t *time, int trylock)
2787 struct hrtimer_sleeper timeout, *to;
2788 struct task_struct *exiting = NULL;
2789 struct rt_mutex_waiter rt_waiter;
2790 struct futex_hash_bucket *hb;
2791 struct futex_q q = futex_q_init;
2794 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2797 if (refill_pi_state_cache())
2800 to = futex_setup_timer(time, &timeout, FLAGS_CLOCKRT, 0);
2803 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &q.key, FUTEX_WRITE);
2804 if (unlikely(ret != 0))
2808 hb = queue_lock(&q);
2810 ret = futex_lock_pi_atomic(uaddr, hb, &q.key, &q.pi_state, current,
2812 if (unlikely(ret)) {
2814 * Atomic work succeeded and we got the lock,
2815 * or failed. Either way, we do _not_ block.
2819 /* We got the lock. */
2821 goto out_unlock_put_key;
2827 * Two reasons for this:
2828 * - EBUSY: Task is exiting and we just wait for the
2830 * - EAGAIN: The user space value changed.
2834 * Handle the case where the owner is in the middle of
2835 * exiting. Wait for the exit to complete otherwise
2836 * this task might loop forever, aka. live lock.
2838 wait_for_owner_exiting(ret, exiting);
2842 goto out_unlock_put_key;
2846 WARN_ON(!q.pi_state);
2849 * Only actually queue now that the atomic ops are done:
2854 ret = rt_mutex_futex_trylock(&q.pi_state->pi_mutex);
2855 /* Fixup the trylock return value: */
2856 ret = ret ? 0 : -EWOULDBLOCK;
2860 rt_mutex_init_waiter(&rt_waiter);
2863 * On PREEMPT_RT_FULL, when hb->lock becomes an rt_mutex, we must not
2864 * hold it while doing rt_mutex_start_proxy(), because then it will
2865 * include hb->lock in the blocking chain, even through we'll not in
2866 * fact hold it while blocking. This will lead it to report -EDEADLK
2867 * and BUG when futex_unlock_pi() interleaves with this.
2869 * Therefore acquire wait_lock while holding hb->lock, but drop the
2870 * latter before calling __rt_mutex_start_proxy_lock(). This
2871 * interleaves with futex_unlock_pi() -- which does a similar lock
2872 * handoff -- such that the latter can observe the futex_q::pi_state
2873 * before __rt_mutex_start_proxy_lock() is done.
2875 raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock);
2876 spin_unlock(q.lock_ptr);
2878 * __rt_mutex_start_proxy_lock() unconditionally enqueues the @rt_waiter
2879 * such that futex_unlock_pi() is guaranteed to observe the waiter when
2880 * it sees the futex_q::pi_state.
2882 ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current);
2883 raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock);
2892 hrtimer_sleeper_start_expires(to, HRTIMER_MODE_ABS);
2894 ret = rt_mutex_wait_proxy_lock(&q.pi_state->pi_mutex, to, &rt_waiter);
2897 spin_lock(q.lock_ptr);
2899 * If we failed to acquire the lock (deadlock/signal/timeout), we must
2900 * first acquire the hb->lock before removing the lock from the
2901 * rt_mutex waitqueue, such that we can keep the hb and rt_mutex wait
2904 * In particular; it is important that futex_unlock_pi() can not
2905 * observe this inconsistency.
2907 if (ret && !rt_mutex_cleanup_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter))
2912 * Fixup the pi_state owner and possibly acquire the lock if we
2915 res = fixup_owner(uaddr, &q, !ret);
2917 * If fixup_owner() returned an error, proprogate that. If it acquired
2918 * the lock, clear our -ETIMEDOUT or -EINTR.
2921 ret = (res < 0) ? res : 0;
2923 /* Unqueue and drop the lock */
2932 hrtimer_cancel(&to->timer);
2933 destroy_hrtimer_on_stack(&to->timer);
2935 return ret != -EINTR ? ret : -ERESTARTNOINTR;
2940 ret = fault_in_user_writeable(uaddr);
2944 if (!(flags & FLAGS_SHARED))
2951 * Userspace attempted a TID -> 0 atomic transition, and failed.
2952 * This is the in-kernel slowpath: we look up the PI state (if any),
2953 * and do the rt-mutex unlock.
2955 static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
2957 u32 curval, uval, vpid = task_pid_vnr(current);
2958 union futex_key key = FUTEX_KEY_INIT;
2959 struct futex_hash_bucket *hb;
2960 struct futex_q *top_waiter;
2963 if (!IS_ENABLED(CONFIG_FUTEX_PI))
2967 if (get_user(uval, uaddr))
2970 * We release only a lock we actually own:
2972 if ((uval & FUTEX_TID_MASK) != vpid)
2975 ret = get_futex_key(uaddr, flags & FLAGS_SHARED, &key, FUTEX_WRITE);
2979 hb = hash_futex(&key);
2980 spin_lock(&hb->lock);
2983 * Check waiters first. We do not trust user space values at
2984 * all and we at least want to know if user space fiddled
2985 * with the futex value instead of blindly unlocking.
2987 top_waiter = futex_top_waiter(hb, &key);
2989 struct futex_pi_state *pi_state = top_waiter->pi_state;
2996 * If current does not own the pi_state then the futex is
2997 * inconsistent and user space fiddled with the futex value.
2999 if (pi_state->owner != current)
3002 get_pi_state(pi_state);
3004 * By taking wait_lock while still holding hb->lock, we ensure
3005 * there is no point where we hold neither; and therefore
3006 * wake_futex_pi() must observe a state consistent with what we
3009 * In particular; this forces __rt_mutex_start_proxy() to
3010 * complete such that we're guaranteed to observe the
3011 * rt_waiter. Also see the WARN in wake_futex_pi().
3013 raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock);
3014 spin_unlock(&hb->lock);
3016 /* drops pi_state->pi_mutex.wait_lock */
3017 ret = wake_futex_pi(uaddr, uval, pi_state);
3019 put_pi_state(pi_state);
3022 * Success, we're done! No tricky corner cases.
3027 * The atomic access to the futex value generated a
3028 * pagefault, so retry the user-access and the wakeup:
3033 * A unconditional UNLOCK_PI op raced against a waiter
3034 * setting the FUTEX_WAITERS bit. Try again.
3039 * wake_futex_pi has detected invalid state. Tell user
3046 * We have no kernel internal state, i.e. no waiters in the
3047 * kernel. Waiters which are about to queue themselves are stuck
3048 * on hb->lock. So we can safely ignore them. We do neither
3049 * preserve the WAITERS bit not the OWNER_DIED one. We are the
3052 if ((ret = cmpxchg_futex_value_locked(&curval, uaddr, uval, 0))) {
3053 spin_unlock(&hb->lock);
3068 * If uval has changed, let user space handle it.
3070 ret = (curval == uval) ? 0 : -EAGAIN;
3073 spin_unlock(&hb->lock);
3083 ret = fault_in_user_writeable(uaddr);
3091 * handle_early_requeue_pi_wakeup() - Detect early wakeup on the initial futex
3092 * @hb: the hash_bucket futex_q was original enqueued on
3093 * @q: the futex_q woken while waiting to be requeued
3094 * @key2: the futex_key of the requeue target futex
3095 * @timeout: the timeout associated with the wait (NULL if none)
3097 * Detect if the task was woken on the initial futex as opposed to the requeue
3098 * target futex. If so, determine if it was a timeout or a signal that caused
3099 * the wakeup and return the appropriate error code to the caller. Must be
3100 * called with the hb lock held.
3103 * - 0 = no early wakeup detected;
3104 * - <0 = -ETIMEDOUT or -ERESTARTNOINTR
3107 int handle_early_requeue_pi_wakeup(struct futex_hash_bucket *hb,
3108 struct futex_q *q, union futex_key *key2,
3109 struct hrtimer_sleeper *timeout)
3114 * With the hb lock held, we avoid races while we process the wakeup.
3115 * We only need to hold hb (and not hb2) to ensure atomicity as the
3116 * wakeup code can't change q.key from uaddr to uaddr2 if we hold hb.
3117 * It can't be requeued from uaddr2 to something else since we don't
3118 * support a PI aware source futex for requeue.
3120 if (!match_futex(&q->key, key2)) {
3121 WARN_ON(q->lock_ptr && (&hb->lock != q->lock_ptr));
3123 * We were woken prior to requeue by a timeout or a signal.
3124 * Unqueue the futex_q and determine which it was.
3126 plist_del(&q->list, &hb->chain);
3129 /* Handle spurious wakeups gracefully */
3131 if (timeout && !timeout->task)
3133 else if (signal_pending(current))
3134 ret = -ERESTARTNOINTR;
3140 * futex_wait_requeue_pi() - Wait on uaddr and take uaddr2
3141 * @uaddr: the futex we initially wait on (non-pi)
3142 * @flags: futex flags (FLAGS_SHARED, FLAGS_CLOCKRT, etc.), they must be
3143 * the same type, no requeueing from private to shared, etc.
3144 * @val: the expected value of uaddr
3145 * @abs_time: absolute timeout
3146 * @bitset: 32 bit wakeup bitset set by userspace, defaults to all
3147 * @uaddr2: the pi futex we will take prior to returning to user-space
3149 * The caller will wait on uaddr and will be requeued by futex_requeue() to
3150 * uaddr2 which must be PI aware and unique from uaddr. Normal wakeup will wake
3151 * on uaddr2 and complete the acquisition of the rt_mutex prior to returning to
3152 * userspace. This ensures the rt_mutex maintains an owner when it has waiters;
3153 * without one, the pi logic would not know which task to boost/deboost, if
3154 * there was a need to.
3156 * We call schedule in futex_wait_queue_me() when we enqueue and return there
3157 * via the following--
3158 * 1) wakeup on uaddr2 after an atomic lock acquisition by futex_requeue()
3159 * 2) wakeup on uaddr2 after a requeue
3163 * If 3, cleanup and return -ERESTARTNOINTR.
3165 * If 2, we may then block on trying to take the rt_mutex and return via:
3166 * 5) successful lock
3169 * 8) other lock acquisition failure
3171 * If 6, return -EWOULDBLOCK (restarting the syscall would do the same).
3173 * If 4 or 7, we cleanup and return with -ETIMEDOUT.
3179 static int futex_wait_requeue_pi(u32 __user *uaddr, unsigned int flags,
3180 u32 val, ktime_t *abs_time, u32 bitset,
3183 struct hrtimer_sleeper timeout, *to;
3184 struct rt_mutex_waiter rt_waiter;
3185 struct futex_hash_bucket *hb;
3186 union futex_key key2 = FUTEX_KEY_INIT;
3187 struct futex_q q = futex_q_init;
3190 if (!IS_ENABLED(CONFIG_FUTEX_PI))
3193 if (uaddr == uaddr2)
3199 to = futex_setup_timer(abs_time, &timeout, flags,
3200 current->timer_slack_ns);
3203 * The waiter is allocated on our stack, manipulated by the requeue
3204 * code while we sleep on uaddr.
3206 rt_mutex_init_waiter(&rt_waiter);
3208 ret = get_futex_key(uaddr2, flags & FLAGS_SHARED, &key2, FUTEX_WRITE);
3209 if (unlikely(ret != 0))
3213 q.rt_waiter = &rt_waiter;
3214 q.requeue_pi_key = &key2;
3217 * Prepare to wait on uaddr. On success, increments q.key (key1) ref
3220 ret = futex_wait_setup(uaddr, val, flags, &q, &hb);
3225 * The check above which compares uaddrs is not sufficient for
3226 * shared futexes. We need to compare the keys:
3228 if (match_futex(&q.key, &key2)) {
3234 /* Queue the futex_q, drop the hb lock, wait for wakeup. */
3235 futex_wait_queue_me(hb, &q, to);
3237 spin_lock(&hb->lock);
3238 ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
3239 spin_unlock(&hb->lock);
3244 * In order for us to be here, we know our q.key == key2, and since
3245 * we took the hb->lock above, we also know that futex_requeue() has
3246 * completed and we no longer have to concern ourselves with a wakeup
3247 * race with the atomic proxy lock acquisition by the requeue code. The
3248 * futex_requeue dropped our key1 reference and incremented our key2
3252 /* Check if the requeue code acquired the second futex for us. */
3255 * Got the lock. We might not be the anticipated owner if we
3256 * did a lock-steal - fix up the PI-state in that case.
3258 if (q.pi_state && (q.pi_state->owner != current)) {
3259 spin_lock(q.lock_ptr);
3260 ret = fixup_pi_state_owner(uaddr2, &q, current);
3262 * Drop the reference to the pi state which
3263 * the requeue_pi() code acquired for us.
3265 put_pi_state(q.pi_state);
3266 spin_unlock(q.lock_ptr);
3268 * Adjust the return value. It's either -EFAULT or
3269 * success (1) but the caller expects 0 for success.
3271 ret = ret < 0 ? ret : 0;
3274 struct rt_mutex *pi_mutex;
3277 * We have been woken up by futex_unlock_pi(), a timeout, or a
3278 * signal. futex_unlock_pi() will not destroy the lock_ptr nor
3281 WARN_ON(!q.pi_state);
3282 pi_mutex = &q.pi_state->pi_mutex;
3283 ret = rt_mutex_wait_proxy_lock(pi_mutex, to, &rt_waiter);
3285 spin_lock(q.lock_ptr);
3286 if (ret && !rt_mutex_cleanup_proxy_lock(pi_mutex, &rt_waiter))
3289 debug_rt_mutex_free_waiter(&rt_waiter);
3291 * Fixup the pi_state owner and possibly acquire the lock if we
3294 res = fixup_owner(uaddr2, &q, !ret);
3296 * If fixup_owner() returned an error, proprogate that. If it
3297 * acquired the lock, clear -ETIMEDOUT or -EINTR.
3300 ret = (res < 0) ? res : 0;
3302 /* Unqueue and drop the lock. */
3306 if (ret == -EINTR) {
3308 * We've already been requeued, but cannot restart by calling
3309 * futex_lock_pi() directly. We could restart this syscall, but
3310 * it would detect that the user space "val" changed and return
3311 * -EWOULDBLOCK. Save the overhead of the restart and return
3312 * -EWOULDBLOCK directly.
3319 hrtimer_cancel(&to->timer);
3320 destroy_hrtimer_on_stack(&to->timer);
3326 * Support for robust futexes: the kernel cleans up held futexes at
3329 * Implementation: user-space maintains a per-thread list of locks it
3330 * is holding. Upon do_exit(), the kernel carefully walks this list,
3331 * and marks all locks that are owned by this thread with the
3332 * FUTEX_OWNER_DIED bit, and wakes up a waiter (if any). The list is
3333 * always manipulated with the lock held, so the list is private and
3334 * per-thread. Userspace also maintains a per-thread 'list_op_pending'
3335 * field, to allow the kernel to clean up if the thread dies after
3336 * acquiring the lock, but just before it could have added itself to
3337 * the list. There can only be one such pending lock.
3341 * sys_set_robust_list() - Set the robust-futex list head of a task
3342 * @head: pointer to the list-head
3343 * @len: length of the list-head, as userspace expects
3345 SYSCALL_DEFINE2(set_robust_list, struct robust_list_head __user *, head,
3348 if (!futex_cmpxchg_enabled)
3351 * The kernel knows only one size for now:
3353 if (unlikely(len != sizeof(*head)))
3356 current->robust_list = head;
3362 * sys_get_robust_list() - Get the robust-futex list head of a task
3363 * @pid: pid of the process [zero for current task]
3364 * @head_ptr: pointer to a list-head pointer, the kernel fills it in
3365 * @len_ptr: pointer to a length field, the kernel fills in the header size
3367 SYSCALL_DEFINE3(get_robust_list, int, pid,
3368 struct robust_list_head __user * __user *, head_ptr,
3369 size_t __user *, len_ptr)
3371 struct robust_list_head __user *head;
3373 struct task_struct *p;
3375 if (!futex_cmpxchg_enabled)
3384 p = find_task_by_vpid(pid);
3390 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3393 head = p->robust_list;
3396 if (put_user(sizeof(*head), len_ptr))
3398 return put_user(head, head_ptr);
3406 /* Constants for the pending_op argument of handle_futex_death */
3407 #define HANDLE_DEATH_PENDING true
3408 #define HANDLE_DEATH_LIST false
3411 * Process a futex-list entry, check whether it's owned by the
3412 * dying task, and do notification if so:
3414 static int handle_futex_death(u32 __user *uaddr, struct task_struct *curr,
3415 bool pi, bool pending_op)
3417 u32 uval, nval, mval;
3421 /* Futex address must be 32bit aligned */
3422 if ((((unsigned long)uaddr) % sizeof(*uaddr)) != 0)
3426 if (get_user(uval, uaddr))
3430 * Special case for regular (non PI) futexes. The unlock path in
3431 * user space has two race scenarios:
3433 * 1. The unlock path releases the user space futex value and
3434 * before it can execute the futex() syscall to wake up
3435 * waiters it is killed.
3437 * 2. A woken up waiter is killed before it can acquire the
3438 * futex in user space.
3440 * In the second case, the wake up notification could be generated
3441 * by the unlock path in user space after setting the futex value
3442 * to zero or by the kernel after setting the OWNER_DIED bit below.
3444 * In both cases the TID validation below prevents a wakeup of
3445 * potential waiters which can cause these waiters to block
3448 * In both cases the following conditions are met:
3450 * 1) task->robust_list->list_op_pending != NULL
3451 * @pending_op == true
3452 * 2) The owner part of user space futex value == 0
3453 * 3) Regular futex: @pi == false
3455 * If these conditions are met, it is safe to attempt waking up a
3456 * potential waiter without touching the user space futex value and
3457 * trying to set the OWNER_DIED bit. If the futex value is zero,
3458 * the rest of the user space mutex state is consistent, so a woken
3459 * waiter will just take over the uncontended futex. Setting the
3460 * OWNER_DIED bit would create inconsistent state and malfunction
3461 * of the user space owner died handling. Otherwise, the OWNER_DIED
3462 * bit is already set, and the woken waiter is expected to deal with
3465 owner = uval & FUTEX_TID_MASK;
3467 if (pending_op && !pi && !owner) {
3468 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3472 if (owner != task_pid_vnr(curr))
3476 * Ok, this dying thread is truly holding a futex
3477 * of interest. Set the OWNER_DIED bit atomically
3478 * via cmpxchg, and if the value had FUTEX_WAITERS
3479 * set, wake up a waiter (if any). (We have to do a
3480 * futex_wake() even if OWNER_DIED is already set -
3481 * to handle the rare but possible case of recursive
3482 * thread-death.) The rest of the cleanup is done in
3485 mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
3488 * We are not holding a lock here, but we want to have
3489 * the pagefault_disable/enable() protection because
3490 * we want to handle the fault gracefully. If the
3491 * access fails we try to fault in the futex with R/W
3492 * verification via get_user_pages. get_user() above
3493 * does not guarantee R/W access. If that fails we
3494 * give up and leave the futex locked.
3496 if ((err = cmpxchg_futex_value_locked(&nval, uaddr, uval, mval))) {
3499 if (fault_in_user_writeable(uaddr))
3517 * Wake robust non-PI futexes here. The wakeup of
3518 * PI futexes happens in exit_pi_state():
3520 if (!pi && (uval & FUTEX_WAITERS))
3521 futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
3527 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3529 static inline int fetch_robust_entry(struct robust_list __user **entry,
3530 struct robust_list __user * __user *head,
3533 unsigned long uentry;
3535 if (get_user(uentry, (unsigned long __user *)head))
3538 *entry = (void __user *)(uentry & ~1UL);
3545 * Walk curr->robust_list (very carefully, it's a userspace list!)
3546 * and mark any locks found there dead, and notify any waiters.
3548 * We silently return on any sign of list-walking problem.
3550 static void exit_robust_list(struct task_struct *curr)
3552 struct robust_list_head __user *head = curr->robust_list;
3553 struct robust_list __user *entry, *next_entry, *pending;
3554 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3555 unsigned int next_pi;
3556 unsigned long futex_offset;
3559 if (!futex_cmpxchg_enabled)
3563 * Fetch the list head (which was registered earlier, via
3564 * sys_set_robust_list()):
3566 if (fetch_robust_entry(&entry, &head->list.next, &pi))
3569 * Fetch the relative futex offset:
3571 if (get_user(futex_offset, &head->futex_offset))
3574 * Fetch any possibly pending lock-add first, and handle it
3577 if (fetch_robust_entry(&pending, &head->list_op_pending, &pip))
3580 next_entry = NULL; /* avoid warning with gcc */
3581 while (entry != &head->list) {
3583 * Fetch the next entry in the list before calling
3584 * handle_futex_death:
3586 rc = fetch_robust_entry(&next_entry, &entry->next, &next_pi);
3588 * A pending lock might already be on the list, so
3589 * don't process it twice:
3591 if (entry != pending) {
3592 if (handle_futex_death((void __user *)entry + futex_offset,
3593 curr, pi, HANDLE_DEATH_LIST))
3601 * Avoid excessively long or circular lists:
3610 handle_futex_death((void __user *)pending + futex_offset,
3611 curr, pip, HANDLE_DEATH_PENDING);
3615 static void futex_cleanup(struct task_struct *tsk)
3617 if (unlikely(tsk->robust_list)) {
3618 exit_robust_list(tsk);
3619 tsk->robust_list = NULL;
3622 #ifdef CONFIG_COMPAT
3623 if (unlikely(tsk->compat_robust_list)) {
3624 compat_exit_robust_list(tsk);
3625 tsk->compat_robust_list = NULL;
3629 if (unlikely(!list_empty(&tsk->pi_state_list)))
3630 exit_pi_state_list(tsk);
3634 * futex_exit_recursive - Set the tasks futex state to FUTEX_STATE_DEAD
3635 * @tsk: task to set the state on
3637 * Set the futex exit state of the task lockless. The futex waiter code
3638 * observes that state when a task is exiting and loops until the task has
3639 * actually finished the futex cleanup. The worst case for this is that the
3640 * waiter runs through the wait loop until the state becomes visible.
3642 * This is called from the recursive fault handling path in do_exit().
3644 * This is best effort. Either the futex exit code has run already or
3645 * not. If the OWNER_DIED bit has been set on the futex then the waiter can
3646 * take it over. If not, the problem is pushed back to user space. If the
3647 * futex exit code did not run yet, then an already queued waiter might
3648 * block forever, but there is nothing which can be done about that.
3650 void futex_exit_recursive(struct task_struct *tsk)
3652 /* If the state is FUTEX_STATE_EXITING then futex_exit_mutex is held */
3653 if (tsk->futex_state == FUTEX_STATE_EXITING)
3654 mutex_unlock(&tsk->futex_exit_mutex);
3655 tsk->futex_state = FUTEX_STATE_DEAD;
3658 static void futex_cleanup_begin(struct task_struct *tsk)
3661 * Prevent various race issues against a concurrent incoming waiter
3662 * including live locks by forcing the waiter to block on
3663 * tsk->futex_exit_mutex when it observes FUTEX_STATE_EXITING in
3664 * attach_to_pi_owner().
3666 mutex_lock(&tsk->futex_exit_mutex);
3669 * Switch the state to FUTEX_STATE_EXITING under tsk->pi_lock.
3671 * This ensures that all subsequent checks of tsk->futex_state in
3672 * attach_to_pi_owner() must observe FUTEX_STATE_EXITING with
3673 * tsk->pi_lock held.
3675 * It guarantees also that a pi_state which was queued right before
3676 * the state change under tsk->pi_lock by a concurrent waiter must
3677 * be observed in exit_pi_state_list().
3679 raw_spin_lock_irq(&tsk->pi_lock);
3680 tsk->futex_state = FUTEX_STATE_EXITING;
3681 raw_spin_unlock_irq(&tsk->pi_lock);
3684 static void futex_cleanup_end(struct task_struct *tsk, int state)
3687 * Lockless store. The only side effect is that an observer might
3688 * take another loop until it becomes visible.
3690 tsk->futex_state = state;
3692 * Drop the exit protection. This unblocks waiters which observed
3693 * FUTEX_STATE_EXITING to reevaluate the state.
3695 mutex_unlock(&tsk->futex_exit_mutex);
3698 void futex_exec_release(struct task_struct *tsk)
3701 * The state handling is done for consistency, but in the case of
3702 * exec() there is no way to prevent futher damage as the PID stays
3703 * the same. But for the unlikely and arguably buggy case that a
3704 * futex is held on exec(), this provides at least as much state
3705 * consistency protection which is possible.
3707 futex_cleanup_begin(tsk);
3710 * Reset the state to FUTEX_STATE_OK. The task is alive and about
3711 * exec a new binary.
3713 futex_cleanup_end(tsk, FUTEX_STATE_OK);
3716 void futex_exit_release(struct task_struct *tsk)
3718 futex_cleanup_begin(tsk);
3720 futex_cleanup_end(tsk, FUTEX_STATE_DEAD);
3723 long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
3724 u32 __user *uaddr2, u32 val2, u32 val3)
3726 int cmd = op & FUTEX_CMD_MASK;
3727 unsigned int flags = 0;
3729 if (!(op & FUTEX_PRIVATE_FLAG))
3730 flags |= FLAGS_SHARED;
3732 if (op & FUTEX_CLOCK_REALTIME) {
3733 flags |= FLAGS_CLOCKRT;
3734 if (cmd != FUTEX_WAIT_BITSET && cmd != FUTEX_WAIT_REQUEUE_PI)
3740 case FUTEX_UNLOCK_PI:
3741 case FUTEX_TRYLOCK_PI:
3742 case FUTEX_WAIT_REQUEUE_PI:
3743 case FUTEX_CMP_REQUEUE_PI:
3744 if (!futex_cmpxchg_enabled)
3750 val3 = FUTEX_BITSET_MATCH_ANY;
3752 case FUTEX_WAIT_BITSET:
3753 return futex_wait(uaddr, flags, val, timeout, val3);
3755 val3 = FUTEX_BITSET_MATCH_ANY;
3757 case FUTEX_WAKE_BITSET:
3758 return futex_wake(uaddr, flags, val, val3);
3760 return futex_requeue(uaddr, flags, uaddr2, val, val2, NULL, 0);
3761 case FUTEX_CMP_REQUEUE:
3762 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 0);
3764 return futex_wake_op(uaddr, flags, uaddr2, val, val2, val3);
3766 return futex_lock_pi(uaddr, flags, timeout, 0);
3767 case FUTEX_UNLOCK_PI:
3768 return futex_unlock_pi(uaddr, flags);
3769 case FUTEX_TRYLOCK_PI:
3770 return futex_lock_pi(uaddr, flags, NULL, 1);
3771 case FUTEX_WAIT_REQUEUE_PI:
3772 val3 = FUTEX_BITSET_MATCH_ANY;
3773 return futex_wait_requeue_pi(uaddr, flags, val, timeout, val3,
3775 case FUTEX_CMP_REQUEUE_PI:
3776 return futex_requeue(uaddr, flags, uaddr2, val, val2, &val3, 1);
3782 SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
3783 struct __kernel_timespec __user *, utime, u32 __user *, uaddr2,
3786 struct timespec64 ts;
3787 ktime_t t, *tp = NULL;
3789 int cmd = op & FUTEX_CMD_MASK;
3791 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3792 cmd == FUTEX_WAIT_BITSET ||
3793 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3794 if (unlikely(should_fail_futex(!(op & FUTEX_PRIVATE_FLAG))))
3796 if (get_timespec64(&ts, utime))
3798 if (!timespec64_valid(&ts))
3801 t = timespec64_to_ktime(ts);
3802 if (cmd == FUTEX_WAIT)
3803 t = ktime_add_safe(ktime_get(), t);
3804 else if (cmd != FUTEX_LOCK_PI && !(op & FUTEX_CLOCK_REALTIME))
3805 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
3809 * requeue parameter in 'utime' if cmd == FUTEX_*_REQUEUE_*.
3810 * number of waiters to wake in 'utime' if cmd == FUTEX_WAKE_OP.
3812 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
3813 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
3814 val2 = (u32) (unsigned long) utime;
3816 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
3819 #ifdef CONFIG_COMPAT
3821 * Fetch a robust-list pointer. Bit 0 signals PI futexes:
3824 compat_fetch_robust_entry(compat_uptr_t *uentry, struct robust_list __user **entry,
3825 compat_uptr_t __user *head, unsigned int *pi)
3827 if (get_user(*uentry, head))
3830 *entry = compat_ptr((*uentry) & ~1);
3831 *pi = (unsigned int)(*uentry) & 1;
3836 static void __user *futex_uaddr(struct robust_list __user *entry,
3837 compat_long_t futex_offset)
3839 compat_uptr_t base = ptr_to_compat(entry);
3840 void __user *uaddr = compat_ptr(base + futex_offset);
3846 * Walk curr->robust_list (very carefully, it's a userspace list!)
3847 * and mark any locks found there dead, and notify any waiters.
3849 * We silently return on any sign of list-walking problem.
3851 static void compat_exit_robust_list(struct task_struct *curr)
3853 struct compat_robust_list_head __user *head = curr->compat_robust_list;
3854 struct robust_list __user *entry, *next_entry, *pending;
3855 unsigned int limit = ROBUST_LIST_LIMIT, pi, pip;
3856 unsigned int next_pi;
3857 compat_uptr_t uentry, next_uentry, upending;
3858 compat_long_t futex_offset;
3861 if (!futex_cmpxchg_enabled)
3865 * Fetch the list head (which was registered earlier, via
3866 * sys_set_robust_list()):
3868 if (compat_fetch_robust_entry(&uentry, &entry, &head->list.next, &pi))
3871 * Fetch the relative futex offset:
3873 if (get_user(futex_offset, &head->futex_offset))
3876 * Fetch any possibly pending lock-add first, and handle it
3879 if (compat_fetch_robust_entry(&upending, &pending,
3880 &head->list_op_pending, &pip))
3883 next_entry = NULL; /* avoid warning with gcc */
3884 while (entry != (struct robust_list __user *) &head->list) {
3886 * Fetch the next entry in the list before calling
3887 * handle_futex_death:
3889 rc = compat_fetch_robust_entry(&next_uentry, &next_entry,
3890 (compat_uptr_t __user *)&entry->next, &next_pi);
3892 * A pending lock might already be on the list, so
3893 * dont process it twice:
3895 if (entry != pending) {
3896 void __user *uaddr = futex_uaddr(entry, futex_offset);
3898 if (handle_futex_death(uaddr, curr, pi,
3904 uentry = next_uentry;
3908 * Avoid excessively long or circular lists:
3916 void __user *uaddr = futex_uaddr(pending, futex_offset);
3918 handle_futex_death(uaddr, curr, pip, HANDLE_DEATH_PENDING);
3922 COMPAT_SYSCALL_DEFINE2(set_robust_list,
3923 struct compat_robust_list_head __user *, head,
3926 if (!futex_cmpxchg_enabled)
3929 if (unlikely(len != sizeof(*head)))
3932 current->compat_robust_list = head;
3937 COMPAT_SYSCALL_DEFINE3(get_robust_list, int, pid,
3938 compat_uptr_t __user *, head_ptr,
3939 compat_size_t __user *, len_ptr)
3941 struct compat_robust_list_head __user *head;
3943 struct task_struct *p;
3945 if (!futex_cmpxchg_enabled)
3954 p = find_task_by_vpid(pid);
3960 if (!ptrace_may_access(p, PTRACE_MODE_READ_REALCREDS))
3963 head = p->compat_robust_list;
3966 if (put_user(sizeof(*head), len_ptr))
3968 return put_user(ptr_to_compat(head), head_ptr);
3975 #endif /* CONFIG_COMPAT */
3977 #ifdef CONFIG_COMPAT_32BIT_TIME
3978 SYSCALL_DEFINE6(futex_time32, u32 __user *, uaddr, int, op, u32, val,
3979 struct old_timespec32 __user *, utime, u32 __user *, uaddr2,
3982 struct timespec64 ts;
3983 ktime_t t, *tp = NULL;
3985 int cmd = op & FUTEX_CMD_MASK;
3987 if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
3988 cmd == FUTEX_WAIT_BITSET ||
3989 cmd == FUTEX_WAIT_REQUEUE_PI)) {
3990 if (get_old_timespec32(&ts, utime))
3992 if (!timespec64_valid(&ts))
3995 t = timespec64_to_ktime(ts);
3996 if (cmd == FUTEX_WAIT)
3997 t = ktime_add_safe(ktime_get(), t);
3998 else if (cmd != FUTEX_LOCK_PI && !(op & FUTEX_CLOCK_REALTIME))
3999 t = timens_ktime_to_host(CLOCK_MONOTONIC, t);
4002 if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE ||
4003 cmd == FUTEX_CMP_REQUEUE_PI || cmd == FUTEX_WAKE_OP)
4004 val2 = (int) (unsigned long) utime;
4006 return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
4008 #endif /* CONFIG_COMPAT_32BIT_TIME */
4010 static void __init futex_detect_cmpxchg(void)
4012 #ifndef CONFIG_HAVE_FUTEX_CMPXCHG
4016 * This will fail and we want it. Some arch implementations do
4017 * runtime detection of the futex_atomic_cmpxchg_inatomic()
4018 * functionality. We want to know that before we call in any
4019 * of the complex code paths. Also we want to prevent
4020 * registration of robust lists in that case. NULL is
4021 * guaranteed to fault and we get -EFAULT on functional
4022 * implementation, the non-functional ones will return
4025 if (cmpxchg_futex_value_locked(&curval, NULL, 0, 0) == -EFAULT)
4026 futex_cmpxchg_enabled = 1;
4030 static int __init futex_init(void)
4032 unsigned int futex_shift;
4035 #if CONFIG_BASE_SMALL
4036 futex_hashsize = 16;
4038 futex_hashsize = roundup_pow_of_two(256 * num_possible_cpus());
4041 futex_queues = alloc_large_system_hash("futex", sizeof(*futex_queues),
4043 futex_hashsize < 256 ? HASH_SMALL : 0,
4045 futex_hashsize, futex_hashsize);
4046 futex_hashsize = 1UL << futex_shift;
4048 futex_detect_cmpxchg();
4050 for (i = 0; i < futex_hashsize; i++) {
4051 atomic_set(&futex_queues[i].waiters, 0);
4052 plist_head_init(&futex_queues[i].chain);
4053 spin_lock_init(&futex_queues[i].lock);
4058 core_initcall(futex_init);